Actively maintained
Changelog
Feature changes, new endpoints and fixes — plus how we handle breaking changes.
Change policy
New endpointsNo extra charge (you pay per call; the number of endpoints does not change the price)
Response shape changesAnnounced in advance; fields are only added, never removed (unless there is a clear defect)
Retiring a featureAnnounced 30 days ahead
Bug fixesShipped immediately, logged on this page
2026-11-10ReleaseScreening values now accept Chinese names — no need to look up a code firstpool/strategy/eventaccept either a Chinese name or a code.pool=央企andpool=shareholder_central_statereturn exactly the same result (517 stocks, measured).- Names match by exact equality first, then by containment. When several fit, the API does not guess — it returns the candidates so you can pick:
strategy=金叉lists MACD golden cross / KDJ golden cross / RSI golden cross / BIAS golden cross / the MACD cross of the day, 5 in total. - If nothing matches you get
422pointing at the full catalogue via?mode=xxx&list=true(codes, Chinese names and groups). - Code-based calls are completely unaffected: no catalogue lookup, no extra request, byte-identical responses.
2026-11-10ReleaseDedicated page for global futures and FX: 55 futures + 10 FX spot pairs, every code in one table- New
/futures-forexcategory page: the full code list for 7 groups of global futures (metals / energy & chemicals / base metals / grains / equity index / rates / currency futures) plus 10 FX spot pairs — every code verified live; anything outside the list returns nothing. - Field differences documented on the same page: futures bars have 7 columns (including
openInterest), FX bars have 5 and no volume — branch on the instrument, never hard-code the A-share 8. - Boundaries stated up front: no domestic Chinese commodity futures, no domestic index or Treasury futures, no softs, no VIX; futures/FX have no quarterly or yearly bars and do not accept
start/end(usecountfor the last N bars). - Three runnable examples: COMEX gold
fuGCdaily, EURUSDfxEURUSDdaily, WTI crudefuCLintraday.
2026-11-09ReleaseHalf the traffic on screening: single-table output drops 42% of the payload, and filter results can be paged/v1/screenaccepts a newslimflag. On single-table outputtablesduplicatesstructuredverbatim, soslim=truedrops the duplicate — measured on a 300-row result: 117 KB to 67 KB, 42.6% smaller, withdataandstructuredbyte-identical.- Multi-table output (investment calendar / macro) is unaffected: those sections differ from each other, so nothing is trimmed.
offsetnow works in filter mode (mode=filter) too: skip the first N rows and take the next slice, pairing withlimitto page through results. It fetches the full set and slices locally, so rows are dropped but values never change.- The MCP tool
ashare_screensupportsslimas well.
2026-11-09DataFactor library settles at 87: the goodwill factor is gone because it could not be used in screening/v1/factorsnow returns 87 factors (previously 88). The removed one isGoodwill— it was not usable inside a filter expression and errored out when written into a condition, so keeping it in the list only produced queries that could not run.- Category counts updated to match: valuation 6 · market 2 · profitability 34 · balance sheet 32 · cash flow 13 — 87 total.
- Goodwill data is still available from the financial statements endpoint (balance-sheet basis).
filters=goodwill:>0now returns422with the list of usable fields, instead of an unhelpful generic error.
2026-11-08DataK-line long ranges in one call: 3, 5 or full history without chunking it yourselfstart/endon daily/v1/klinenow accept any span: ask for 3 years, 5 years or everything since listing, and a single call returns the complete series — no need to split by year and stitch the results.- One request per symbol instead of many: fewer calls, simpler code, and less work when preparing historical data for backtests.
countis still the upper bound ("how many bars at most"), taken from the most recent N bars, matching the response order (data[0]is the latest bar).- Weekly / monthly / quarterly / yearly bars still use
count(1212 bars already covers 20+ years of monthly bars); unchanged.
2026-11-08ReleaseL2 large-order fills now offer a full-session mode: the whole day in one call, enough to compute net inflow- New
GET /v1/bigorder-flow-fullreturns every fill at or above the threshold for the day, without the ~200-row cap of/v1/bigorder-flow. - Use it to compute full-day net inflow, active-buy vs active-sell totals, and large-order turnover. Computing a full day from the latest 200 rows gives the wrong number.
- Top-level
metareturnstotal,returned,deduped(merged duplicates),pages,complete(whether the whole day was fetched) andsummary(amount totals by side). - Duplicate pushes from the closing auction (15:00:00) are merged automatically;
dedupedis the merged count. - New MCP tool
ashare_bigorder_flow_full; total tools 91 → 92. - Timing: about 20-30 s for an active stock at the 300k CNY threshold, ~10 s at 1M, ~2 s at 3M. Use 1M or above when you only need the conclusion.
- Institutional tier only; 100 calls per day.
2026-11-02EndpointsUsage query now covers every daily-capped endpoint: see how many calls each has left today/v1/usageadds aheavy_quotasfield returning used / limit / left for all 22 daily-capped endpoints (batch fetch, backtests, L2 big orders).- Each entry looks like
"snapshot-batch": {"used": 0, "limit": 1000, "left": 1000}; whenlefthits zero that endpoint returns429withretry_after_s. - Quotas count calls, not symbols, and reset at 00:00 Beijing time — check before starting a batch job.
- The existing
heavyfield (kline-full only) is unchanged; no breaking change. - Every endpoint page now shows a per-endpoint quota block with the per-call cap and daily calls by tier.
2026-11-02GuidesNew guide: batch fetching — 5,000 requests down to 23- Nine batch endpoints return 10–1000 symbols per call: quotes, K-lines, full snapshots, money flow, technicals, chips, shareholders, financials, company profiles.
- Includes the per-call cap and daily quota table, runnable Python (chunked), and the real request count for a full-market scan.
- Explains per-symbol billing, automatic refunds for failed symbols, daily resets, and three pitfalls (A-shares only, prefixes required, exceeding the cap rejects the call).
2026-11-01ReleaseL2 big-order queue adds cancellation stats: cancel_count / cancel_ratio / cancel_volume / cancel_amount/v1/bigorder-queuenow addscancel_count,cancel_ratio(0–1),cancel_volume(lots), andcancel_amount(CNY) to top-levelmeta.- Scope is explicit:
cancel_stats_scope="returned_rows"means the stats cover the returned sample, not the full day;truncatedstill shows whether rows were cut. - Use it to spot posted-then-pulled big orders, fake seals, and unreliable depth; pair with
/v1/bigorder-flowto separate “posted but not filled” from “actually traded”.
2026-11-01ReleaseDouble 11 Institutional annual plan: early bird ¥1,599/year, 10 seats- Institutional annual: Nov 1–5 early bird ¥1,599/year (10 seats); Nov 6–11 ¥1,699/year (20 seats); back to ¥1,990/year on Nov 12.
- Includes all current Institutional benefits for 365 days; roadmap features get priority access, with no promised release dates.
- Annual subscription, not equity or investment. No return guarantee, no auto-renewal.
- Monthly Institutional remains available; contact us to activate annual and issue the key.
2026-11-01ReleaseAnalyst Agent is live: /v1/analyst and a new MCP tool- New
GET /v1/analyst?snapshot_id=...: reads a scanner snapshot and T+1/T+5/T+20 evaluation, returning return, benchmark, excess, hit rate, winners/losers and data gaps. - New MCP tool
ashare_analyst; tool count 90 → 91. - Read-only: no stock picking, no forecasts, no buy/sell recommendations; snapshots and evaluations are unchanged.
- Optional BYOK: with
X-LLM-Keythe model polishes the review; without it, a deterministic review is returned. - When evaluation data is unavailable it says the window is not ready; it never fills zeros or invents returns.
2026-10-31ReleaseWatchlists are live: /v1/watchlist, batch quotes and 4 MCP tools- New
GET /v1/watchlist:action=list|add|remove, with watchlists isolated per API key. - New
GET /v1/watchlist/quotes: latest quotes for a watchlist in one call; charged as one data call. - Four new MCP tools:
ashare_watchlist_list/ashare_watchlist_add/ashare_watchlist_remove/ashare_watchlist_quotes; total tools 86 → 90. - Listing, adding and removing do not consume data calls; only batch quotes count as one data call.
- Tier limits: Trial 1 list / 20 symbols; Standard 3 / 100; Pro 10 / 500; Institutional 50 / 2,000.
- Data is isolated by tenant; rotating the key keeps the same tenant and the watchlist.
2026-10-31ReleaseStock Screener (scanner Agent) is live: /v1/scanner and a new MCP tool- New
GET /v1/scanner: natural-language conditions → read-only tools → picks + citations + tool trace. - New MCP tool
ashare_scannerfor Claude Code / Cursor / Codex and other clients. - BYOK: besides the ashareapi key, send
X-LLM-Key(remote MCP:ASHARE_LLM_KEY); model tokens are paid by the customer. - Scanner quota is separate: Pro ¥99 includes 2,000 lifetime calls; Institutional ¥199 includes 10,000/month. It does not consume the 500k data-API pool.
- Only
ok:truewith non-empty picks is charged; failures and empty results are refunded. - ⚠️ Streaming is not supported yet; not investment advice, no automated trading.
2026-10-31GuidesAdded the Personal Agent hub plus Today / Grok Bot / Muse / Dots / Duet setup guides- Hub:
/docs/guides/personal-agentexplains what a personal agent is, which clients support MCP, and how each entry works. - One endpoint:
https://api.ashareapi.com/mcpwithAuthorization: Bearer <ashareapi Key>. - Today: Settings → Connected apps → Add custom MCP → Add via link → Bearer token (verified).
- Grok Bot: add a remote MCP under Connect apps with a Bearer header; team workspaces follow the Cursor connector policy.
- Muse: create a Custom Connector under Settings → Connectors. Dots: add the MCP to the user’s ChatGPT connectors/plugins. Duet: add it to the user’s Claude connectors.
- Users bring their own accounts and subscriptions. We provide the MCP service, keys and tutorials, not Today/ChatGPT/Claude/Grok/Muse accounts.
2026-10-30FixFixed missing data and codes filtering; added per-symbol status, risk flags and macro history depth- Backfilled the full-market daily bars for 2026-10-29 (5,569 symbols):
/v1/market-klineand/v1/kline-fullreturn that session again. If a session lags, the sync now refills and retries, and reports "not synced" instead of silently skipping the whole day. /v1/market-klinenow applies thecodesfilter before paging — previouslycodes=sh600519&limit=1paged first (taking the lowest-sorting symbol) and filtered afterwards, so the requested symbol came back empty.600519andsh600519are equivalent./v1/market-klinewithcodesnow returnsrequested/returned/missing: which symbols were not returned and why (not_covered= outside this endpoint's scope;no_data= no bar for that session). When paging truncates the result it is flaggedpartial:true— a symbol absent from this page does not mean it has no data./v1/macrogainedlimit(default 12, max 120): it previously returned a fixed 3 rows. For multi-year LPR / CPI / PPI trends passlimit=60(LPR yields 58 rows, measured)./v1/screenresults now carry a风险标记(risk flag) column (-/ST/退市整理), andstructuredrows also carryisST/isDelisting/status; all five modes are covered. Flags only, no implicit filtering — you decide whether to exclude them, so backtests stay reproducible.- ⚠️ All 9 batch endpoints now return
ok:falsewhen nothing was retrieved (previously alwaysok:true). Partial success is stillok:true; the per-symbolok/errorfields are unchanged. If your code uses the outerokalone to decide the whole batch succeeded, also checkreturned. - When the source rate-limits, per-symbol
errornow says "data source rate-limited, retry in about 60 seconds" instead of lumping it in with "suspended / invalid code". - When the source rejects a parameter value (e.g. a non-existent
typeonscreen), the response isok:false+reason:"invalid_param"+hint(with the source's own message and the right way to list valid values:?mode=xxx&list=true), and the call is not billed. - ⚠️
/v1/consensusnow returnsnullforinstitutionCnt: that column carries no usable value at the source, and a constant0was being read as "no analyst coverage". Do not use it to judge coverage;targetPriceand the per-year estimates are unaffected. /v1/financedocuments the bank reporting basis: commercial banks do not provideTotalAssets— useTotalLiability + TotalShareholderEquityfor total assets andDebtAssetsRatiodirectly for the debt ratio; do not derive it from current + non-current assets.
2026-10-29GuidesNew setup guides for Qoder, Trae and WorkBuddy- Qoder: edit
~/.qoder-cn/mcp.jsonwith the URL and headers — custom headers are supported and the config reloads automatically. The manual-add dialog only offers a name and service address, so paid tools require the config file. - Trae: Settings → MCP → Add manually, then paste a JSON block. Trae HTTP configs carry a
headersfield, and per-project.trae/mcp.jsonworks once "Enable project-level MCP" is on. - WorkBuddy: 专家·技能·连接器 → Connectors → Custom connector → Add MCP, then click Trust on the connector management page — it is not active until you do.
- All three accept custom request headers, so paid tools work. The China platform overview now covers nine platforms.
2026-10-28ReleaseStock screener (scanner Agent): ashare/scanner@deepseek-flash- Available from Pro (¥99). The model picks read-only tools from the question (limit-up / broken-board / screen / quote / market overview); arguments are validated by the harness.
- Supports first-board + fund flow, broken-board + fund flow, factors, labels, rankings, strategies and events; responses include
x_citationsandx_tool_trace. - ⚠️ Conditional stock screening only: no price prediction or trading advice; tool failures degrade; Research Agent Team is still planned and will be announced separately.
2026-10-27ReleaseOpenAI-compatible endpoints: /v1/chat/completions + /v1/responses- Set
base_urltohttps://api.ashareapi.com/v1andapiKeytoct-xxx:sk-yyy(our key, then your model key) — works with the OpenAI SDK, Cherry Studio, ChatBox and OpenWebUI. modelsupportsashare/deepseek-flash/gpt-5.6-luna/claude-*and similar;stream=trueis streaming SSE, whilestream=falsereturns the full JSON.- ⚠️
toolsis ignored in v1; Responses does not support server-sideprevious_response_id. Tier: Pro (¥99) and above. /v1/modelsis live too: the realdeepseek-flash/gpt-5.6-luna/claude-*list.
2026-10-27ReleaseAnthropic Messages protocol: /v1/messages- Anthropic Messages-compatible endpoint for Claude Code / Anthropic SDK: set
x-api-keytoct-xxx:sk-ant-yyy(our key, then your Anthropic key);Authorization: Bearer ct-xxx+x-api-key: sk-ant-yyyalso works. max_tokensis required; SSE is supported (streaming: upstream tokens are forwarded as they are generated;stream=falsereturns the full message);claude-*withsk-ant-...routes to Anthropic, whiledeepseek-*/claude-*withsk-...routes to DeepSeek.- ⚠️
toolsis ignored in v1 (normal chat works; Claude Code file/terminal tools do not); for OpenAI use/v1/chat/completions//v1/responses. Tier: Pro (¥99) and above.
2026-10-27Data/v1/fund-flow now supports a Shanghai/Shenzhen history range (start / end)- Pass
start+end(YYYY-MM-DD, single code) to get daily money flow over that range; omit both to keep the latest-day row. - History ranges are Shanghai/Shenzhen only (
sh*/sz*); HK returns the latest day only. ⚠️ US stocks have no money-flow data (every tier returns an empty result) — for US, use/v1/fund-short(short selling). Amounts are raw CNY/HKD units. - The range and the latest-day row can differ slightly because of upstream conventions — use one channel per request; do not mix with the 5/10/20-day cumulative fields from
/v1/fund.
2026-10-26Endpoints/v1/auction adds is_limit_down — the "touched limit-down during auction" flag- The exact mirror of the existing
is_limit_up(touched limit-up):is_limit_down= auction low ≤ limit-down price, returned at the top level undermeta, for bothmode=stockandmode=stock-history. - The limit-down price is taken from a raw source field (not computed by us), so main board 10% / ChiNext·STAR 20% / ST 5% / Beijing Exchange 30% are correct by construction — no formula needed on your side.
metaalso returnslprice(auction low; derived from second-level ticks forstock-history), so you can verify the flag yourself.- The flag only answers "did the auction reach the limit-down price" — it does not mean the price stayed locked there, and says nothing about the session after the open.
2026-10-26ReleaseMCP now accepts the key as a URL parameter (?key=)- For clients that cannot set custom request headers:
https://api.ashareapi.com/mcp?key=YOUR_KEYworks directly — the typical case is a ChatGPT web custom connector. - Tool coverage, quota, rate limits and billing are identical to the
Authorization: Bearermethod — same key, same quota, no extra consumption. - If both are present, the header wins. For production we still recommend the header: the key stays out of the URL.
2026-10-25EndpointsNew: /v1/stocks — the full A-share universe in one call- Returns every currently listed A-share (~5,576) in one request: code, name, market, listing date, industry, sector, issue price and registered capital — ready to seed a stock pool or database.
- How it differs from
/v1/search: search requires a keyword and returns one match; this endpoint needs no keyword and returns the whole list. - Filter with
market=sh|sz|bjorindustry=..., paginate withlimit/offset. The full response is about 1 MB — pull it once at setup, then fetch incrementally. - Coverage is currently listed stocks: delisted names are excluded, and new listings are added automatically at call time.
- Available from the Trial tier (¥9.9).
2026-10-24UpdateThe ¥199 tier is now called "Institutional"- The former "Unlimited" tier is now Institutional. It is positioned around L2 microstructure data (large orders / seat-level detail / market-wide capital ranking), not around a call-count package.
- Entitlements are unchanged: unlimited calls, 50 IPs, all 90 endpoints, a dedicated quota for heavy requests. Existing subscribers keep their key — nothing to re-issue.
- Why the rename: this layer is called "Level-2 enhanced" in brokerage terminals and is usually billed as an institutional tier. The name should say what it is, not how many calls it includes.
2026-10-24EndpointsNew: 6 short-term sentiment and institutional-view endpoints/v1/market-mood: break rate plus a 9-trading-day series of success and profit rates, for judging how hot or cold the short-term tape is./v1/limit-reason: the full reason text behind every past limit-up of one stock, with its sector and its "dragon rank of the day"./v1/broken-board: stocks whose limit-up was smashed open on a given day, with theme tags. Compare with the limit-up pool to see that day's seal quality./v1/tail-chase: about 292 stocks market-wide with grab amount, theme and change. A few symbols are withheld by the data source and carry amaskedflag; amounts are complete./v1/northbound: foreign holdings by A-share sector, published quarterly, with net flow and total holdings. Pass a quarter-end date such as2026-09-30./v1/new-high: new-high counts aggregated by theme — where the new highs cluster is where the money is working.- All 6 endpoints are part of the Institutional (¥199) tier.
2026-10-24EndpointsNew institutional participation board /v1/lhb-agency (Unlimited tier)/v1/lhb-agencyreturns every stock that institutional seats traded on the Dragon-Tiger list on a given day, in one call — 51 stocks on the measured session. It answers "what did institutions buy today" (whole market, filtering upwards); for "who bought this stock" (one stock, looking down into branches) use/v1/lhb-seats.net_buyis institutional net buying in CNY: buys minus sells, and it can be negative — 26 of the 51 measured rows were negative (institutions net selling). Do not take the absolute value as buying volume.- Each row also carries: seats involved
join_num, change, amplitude, turnover rate, float market cap, total market cap, and the theme plates it belongs to. - ⚠️ The Dragon-Tiger list is post-close data: leaving
dateempty gives the data source's latest trading day — trust thedatein the response. - Parameters:
date(YYYY-MM-DD) ·sort(net buy / seats / change / turnover …) ·limit.
2026-10-24DataQuant capabilities consolidated into Pro- Backtesting, factor library (88), strategy list (29), methodology playbooks (25) and factor screening are now Pro (¥99) only
- Existing Standard subscribers keep full access — the above remain available for the quota you already purchased (no expiry date), at your original quota and rate limits
- Standard (¥29.9) keeps all data endpoints + 35-field snapshot + full-history K-line; adds notices · news · limit-up · suspensions · buybacks · disclosure calendar
- ⚠️ From 2026-10-24, new Standard purchases follow the new scope; choose Pro if you need quant capabilities
2026-10-23EndpointsNew stock scores and analyst reports endpoints/v1/scorestock scores (Pro+ · A-shares only): five dimensions — basic / technical / composite / fund / risk — each with week / month / quarter changes. Measuredsh600519: composite 79.73, basic 85.66, technical 72.74, fund 78.75, risk 82.56.codesaccepts a comma list for side-by-side comparison.- ⭐ Structured output: every dimension is
{value, chg_week, chg_month, chg_quarter}— no string parsing needed (upstream returns mixed text such as85.66 (周↓-1.13 月↓-0.84 季↑+85.66)). ⚠️chg_quarteris the cumulative change this quarter (scores reset quarterly, so it usually equalsvalue), not zero. /v1/reportanalyst reports (Pro+ · A-shares and HK): withoutidyou get a list (title / broker / rating / type / time); withidyou get the full text as Markdown (measured 2,254 characters for a Tencent report, including financial commentary).- ⚠️ At most 20 rows per list call (
limitabove 20 returns 422); page withoffset; the "N total" figure is the historical count, not the page size. No date-range filter — finding a specific period means paging. - ⚠️
/v1/scoreis A-shares only (HK/US return 422);datereaches back to late April 2026 (earlier dates are empty, not a failure). - Tiers: both scores and reports are Pro (¥99) — they are opinion-type / content-type high-value data (scores sit alongside rating / consensus; reports are broker views plus full text), so they are not opened up to the Standard tier.
2026-10-23EndpointsUS market live: 16 entry points + three major indices (S&P 500 is usINX)- The US market now has 16 entry points: quotes / K-lines / intraday / technicals / financial statements / dividends / profile / analyst ratings / earnings disclosure dates / trading halts / IPOs / search / news / filings / short selling / US screening.
- ⭐ The three major indices are queryable (
/v1/quoteand/v1/kline): DowusDJI52,867.98 · NASDAQusIXIC26,008.86 · S&P 500usINX7,663.43 (measured 2026-10-23). ⚠️ Indices aredayonly — no minute bars. - ⚠️ The S&P 500 code is
usINX, notusSPX—usSPX/usGSPC/usSPX500all return nothing. This is the most common trap. /v1/fund-shortcovers US short selling: measuredusAAPLon 2026-09-30 — short ratio 0.83%, days to cover 2.82, shares shorted 121m. ⚠️dateis required, otherwise only identifier fields come back.- US screening works:
/v1/screen?market=us, but the expression field must bePeTTM(case-sensitive —PE_TTMfails). /v1/klinedescription corrected: it previously said HK/US "do not support date ranges"; the actual reason is they are not offered (the same ticker returns different price conventions with and without a range across channels) — usecountfor the last N bars.- ⚠️ Not available for the US (coverage limits): shareholder detail / events calendar / buybacks / consensus estimates / per-stock money flow / index constituents / chip distribution / sectors / LHB / margin trading. Use
/v1/ratingfor US target prices and estimates.
2026-10-23EndpointsHong Kong coverage complete: 22 entry points (10 new endpoints)- Hong Kong now has 22 usable entry points (18 command-level + 4 sub-command-level): quotes / technicals / financials / shareholders / dividends / profile / news & filings / events calendar / IPO / search / buybacks / consensus / trading halts / disclosure calendar / analyst ratings + per-stock money flow / short selling / southbound holdings / index list.
- 10 new endpoints:
/v1/suspensiontrading halts (CN/HK/US) ·/v1/buybackshare buybacks ·/v1/consensussell-side consensus (target price + estimates) ·/v1/ratinganalyst ratings (HK/US only; measuredhk00700average target 732.68 from 49 analysts) ·/v1/disclosureearnings disclosure calendar ·/v1/fund-flowper-stock money flow (A-shares/HK) ·/v1/fund-shortshort selling (HK/US) ·/v1/fund-southsouthbound holdings (quarterly; measured Tencent at 12.03%) ·/v1/index-listindex list (1,223) ·/v1/index-constituentindex constituents (CSI 300 = 300 names). /v1/index-constituentnow covers HK indices:hkHSI(Hang Seng) /hkHSCEI(HSCEI) /hkHSTECH(Hang Seng TECH), 20 names each. ⚠️ Fields differ by market: A-share indices returncode/name/IndexName; HK indices return Chinese keys代码/名称/涨跌幅/成交额(two extra columns). One index per request (comma-separated returns empty)./v1/fundcannot query HK stocks (A-share composite: LHB / block trades / margin) — passinghk*returns 422 pointing to/v1/fund-flow.- Conventions to note:
/v1/fund-shortrequiresdate(without it, and before the day's data is published, only identifier fields are returned);/v1/ratingis HK/US only (A-shares use/v1/consensus);/v1/consensushas no target price for HK (only*Forecastcolumns). - ⚠️ Not available for HK: chip distribution / sectors / hot list / LHB / ETF / macro / futures / FX / convertible bonds / industry chain / research / margin trading.
2026-10-23EndpointsETF upgrades: /v1/etf gains historical snapshots · price adjustment (fq) stays a paid-tier capability/v1/etfnow acceptsdate(historical snapshot): withkind=概览you get that day's premium/discount, size and valuation percentile (measuredsh510300on 2026-06-30: change 1.21%, premium 0.13%, PE 14.54 — clearly different from today's 1.86% / -0.15% / 12.78); withkind=持仓you get the holdings of that day. Omittingdatekeeps the previous output byte-for-byte.- Price adjustment (
fq) remains a paid-tier capability — free K-lines are fixed to forward-adjusted:/v1/klinedoes not exposefq(passing it returns 422 with guidance), so a given code always has one price convention on the free endpoint — no mixing two adjustment modes from the same endpoint. Fornofq(real traded prices) orhfq(back-adjusted, for return attribution) use/v1/kline-full(Standard, A-shares only) or/v1/kline-batch(Pro). - Clarified capability scope:
/v1/chip(chip distribution) works for ETFs too (measured: average cost 4.39, profit ratio 13.47%), previously described as "A-shares only";/v1/margin-tradenow states ETFs are not margin-trading eligible, soNOT_DISCLOSEDshould not be retried as if it were a data delay. - Also:
/v1/etf?kind=持仓is documented as Top 20 holdings (previously described as "full list"; measured 20 rows for bothsh510300andsz159915).
2026-10-23Endpoints/v1/screen now starts at Standard — all five screening modes grouped into Standard- Screening (
/v1/screen) now requires Standard (¥29.9) or above. One endpoint, five modes: factor screening (88 factors in any combination) / attribute labels (108, incl. SOEs, ST, below-book, recent IPOs) / market-wide rankings (composite, capital, fundamentals, risk and technical scores) / strategy screening (98: MACD golden cross, 200-day breakout, high dividend, master strategies…) / event screening (42: dividends, earnings pre-announcements, private placements, lock-up expiry, insider changes…). - Why tier by capability rather than by volume: screening returns a research-ready result set in a single call — it is the entry point of a research workflow. Capability-type endpoints are tiered by plan, while volume-type endpoints are metered by call count.
- Unaffected on Trial: quotes, K-lines, hot lists, market overview and the gainers/losers distribution remain available.
- What Standard adds: full-history K-lines (
/v1/kline-full), announcements (16 years), 7×24 news flashes, financials, backtesting, plus the factor and methodology libraries. - ⚠️ If screening is your main workload, Standard also raises the quota to 100,000 calls (Trial: 10,000) and the rate limit to 120/min (Trial: 30/min).
2026-10-22Endpoints/v1/limit-up gains mode=auction — auction-sealed limit-ups (sealed at the 09:25 fix)- Find stocks that sealed the limit at the call auction — the first table short-term traders check before the open.
mode=auctionkeeps only stocks whose 09:25 fixing price equals the limit price (one-word boards / aggressive auction bidding); everything else is dropped. - Definition matters: a stock that touched the limit during the auction but was knocked off before the fixing does not count — those re-seal after the open and report a later
limit_time(e.g. 09:30). The table is "sealed at fixing", not "touched the limit". - One call replaces per-stock lookups: instead of asking "did this one seal at the auction?" one symbol at a time, get the whole market list in one request, with
reason/seal_amount/streak/main_net. - Accuracy: cross-checked against the independent "fixing price == limit price" definition over 3 trading days / 147 limit-up stocks — zero false positives, zero misses.
- Tier unchanged:
/v1/limit-upstays on Standard (¥29.9) and above;mode=auctionis available at the same tier, no gate raised.
2026-10-22Endpointsashare_agent_search extended: WeChat / web search + sentence-level location- WeChat article search:
scope=wechatsearches inside WeChat Official Accounts (equivalent to restricting tomp.weixin.qq.com);scope=websearches the open web. This covers material outside our own database — industry policy, macro commentary, non-A-share companies. - Sentence-level location: web-text evidence now carries
start_char/end_char(character span), so you can answer *which sentence, which character* — not just a link.end_charmatches the length ofvalue, so it reconciles. - Relevance score: each web / WeChat result carries
rerank_score(0–3, ≥1.5 is good) andquality(good/fair) — use it to judge whether the batch is good enough, not just the count. - Optional replay reconciliation:
replay=trueruns one extra pass and compares the output, markingmatch/non_deterministic/ etc. ⚠️ On the model path it is usuallynon_deterministic(same input, different output bytes) — this does not mean the conclusion is unreliable. Off by default. - ⚠️ Bring your own model key: web / WeChat search uses
X-LLM-Key(cost is on your key); without it we do not fall back to our own quota. - ⚠️ With an explicit
scope=web/wechat, only that channel is searched — if nothing is found you get an honestok:falsewith the reason, never a silent substitute from another channel.
2026-10-21EndpointsL2 main-capital trio: historical big orders / four-way split / whole-market ranking (Unlimited)/v1/bigorder-history— historical big-order net flow: up to 100 trading days back, one value per day. This closes the biggest gap ("big orders only cover today") — tell whether the big money is accumulating steadily or just spiking today, and on which day it turned./v1/bigorder-split— four-way main-capital split: minute-level, main buying split into active buy + passive buy and main selling into active sell + passive sell, plus the volume-weighted average price of main buying / selling. High active buy = someone is sweeping the offer; high passive buy = your bids are being hit. Historical lookup supported, no row cap./v1/bigorder-rank— whole-market main-capital leaderboard: all ~5,568 stocks in one call with main buy / sell / net, 8 sort modes (net in / net out / main buy / main sell / gain / loss / turnover / market cap), pagination and historical lookup. The fastest answer to "what is the big money buying today".- Self-checkable identities:
main_buy + main_sell = main_netandmain_buy = active_buy + passive_buy— hold for 100% of all 5,568 rows, so you can verify the data yourself. - How they fit together: start from
bigorder-rankto pick stocks across the whole market, then drill down withbigorder(minute net flow) /bigorder-split(four-way structure) /bigorder-flow(tick trades) /bigorder-queue(orders and cancellations). Six endpoints form a complete L2 main-capital chain. - ⚠️ Unlimited tier only: Pro and below receive a clear message with an upgrade link.
2026-10-21Endpoints/v1/auction moves to the Pro tier (¥99 and above)- Why: call auction is the front end of the limit-up chain.
mode=stock-history(second-level ticks) needs 2 data-source calls per request and shares one data channel with limit-up and review (serialized, minimum 0.35s interval) — heavy use on the Standard tier would degrade real-time performance for other endpoints on the same tier, so auction now sits in Pro. - Who can use it: Pro (¥99) and above (Unlimited included). Standard / Trial keys receive a clear explanation plus an upgrade link.
- Unchanged:
/v1/minute,/v1/limit-upand/v1/reviewstay on Standard (¥29.9) and above. - Standard users: upgrade at the price difference to unlock it; existing quota and call history are unaffected.
- WARNING Breaking change: Standard keys now get 403 on
/v1/auction(the response includesmin_tierand an upgrade link).
2026-10-20EndpointsNew MCP tool ashare_agent_search: one-question data fetch + evidence chain (60 → 61 tools)- Turns a natural-language question straight into data: "how is the money flow on CATL lately?" returns the data plus full evidence, with no need to guess which endpoint to call.
- Every result is verifiable: carries
value_hash/source_endpoint/params/as_of, enough to independently re-fetch and compare. - Per-claim verdicts:
verification.claims[]states explicitly which claims are verified and which are not — the main difference from comparable products. - WARNING Coverage boundary: supports one or more A-share symbols for money flow / quotes / filings & news / Dragon-Tiger list / block trades / margin trading (e.g. "which of CATL and BYD has more money inflow"); sectors, market-wide stats, technicals, fundamentals and predictions are explicitly refused rather than answered with a different metric.
2026-10-20EndpointsScreening, expanded: one endpoint, five ways to screen (tags / ranking / strategies / events)- Stock screening now takes a
modeparameter; factor screening is unchanged and remains the default (mode=condition). mode=label— screen by attribute tag: central SOEs / SOEs / private / QFII, ST / delisting / broken IPO, below-book, loss-making, cash-flow structure, listing age, market-cap band … 108 tags;asset=etfswitches to ETF theme pools.mode=ranking— whole-market ranking: composite / capital / fundamentals / risk / technical scores, margin balance and its daily-weekly-monthly-quarterly-yearly change, main and retail fund flows, northbound activity, limit-up streak …mode=strategy— technical and style strategies: MACD golden cross, KDJ golden cross, break above the 250-day line, first limit-up, consecutive limit-ups, high dividend, master strategies (Buffett / Graham / Fisher …) — 98 strategies.mode=event— event-driven: dividend plan/resolution/implementation/ex-date, earnings pre-announcement/express/release, private placement before and after listing, management changes and insider trading, block trades, lockup releases … 42 events.- ⭐ Not sure which values exist? Pass
list=true— it returns every available value for that mode, with Chinese names and groups. No guessing.
2026-10-20EndpointsNew: trading calendar (market open/closed, last trading day of the year) — free- New trading calendar — for each date, whether the market is open, and whether it is the last trading day of the week / month / quarter / year. Handy for settlement, expiry and rebalancing dates.
- Three ways to query: no params = current month ·
date=2026-10-16= a single day ·start+end= any range, oryear=2026= a whole year. Addtrading_only=truefor market-open days only (~242 rows a year). - ⚠️ Not the same as
/v1/calendar— that one lists per-stock corporate events (dividends, lockup releases, earnings dates); this one is only about whether the market is open. - Free endpoint (no API key required).
2026-10-19ReferenceEmpty-result docs: how to spot a temporarily unavailable data source (the degraded field)- An empty result is not a failure now covers one more signal:
degraded: truein the response means the data source was temporarily unavailable (for example, briefly rate limited) — not that this instrument has no data. - The response also carries
reasonandretry_after_s(how many seconds until you can retry) — retrying shortly usually works, so do not skip it as "no data". - An
ok:falsewithoutdegradedis what really means "this instrument has no data for this dataset"; that meaning is unchanged, and the response shape for normal requests is completely unchanged.
2026-10-18EndpointsNew: multi-source evidence (filings + news, deduped, with hashes)- New multi-source evidence — one query searches CNINFO filings (both exchanges, official first-hand) and Eastmoney news, normalised into a single table.
- Cross-source dedupe: when one story is republished by several outlets it appears once; the rest are grouped under
also_on. - Every row carries evidence hashes:
cit_endpoint/cit_field_path/cit_value_hash/cit_retrieved_at— enough to reproduce and verify the row independently. - Authority tiers: filings
authority=1.0(official disclosure platform), news0.5(portal). Results are quota-balanced per source rather than purely time-sorted, otherwise filings would be buried by same-day news. - Filings link to the original PDF; news links to the article. Results are cached for 5 minutes.
- Who can use it: Standard (¥29.9) and above; one query searches filings and news together, so it bills as 2 calls.
- Add
scope=reportto include broker research: headline + broker + date + a link to the report page. Research gets its own quota slot, so news never crowds it out, and it is not billed extra. Research carries no abstract, sosnippetis empty. - Evidence hashes now use a published algorithm:
cit_value_hashis a 16-hex digest with the algorithm named incit_hash_alg, and the full formula is in the endpoint description.cit_endpointnow points at an address you can actually call; the original data source moved tocit_upstream.
2026-10-18EndpointsNew: 7x24 financial news flash (market-wide)- New news flash — a market-wide news feed (Cailianshe and others), newest first, each item carrying
time/title/content/source/url. - How it differs from per-symbol news: that one needs a stock code and covers one name; this is a market-wide wire — use it for "what happened just now", and per-symbol news for "what is the story on this stock".
- Paging:
indexstarts at 0 with a step equal to your previouslimit; it only pages backwards. limitis capped at 200; results are cached for 1 minute, so repeat requests within a minute do not cost extra.- Who can use it: Standard (¥29.9) and above.
2026-10-18EndpointsNew: L2 large-order queue + dragon-tiger seat detail (Unlimited tier)- New L2 large-order queue — each pending order above a threshold (>=500 lots or >=300k CNY) with limit-price and cancelled flags. It answers whether a limit-up wall is real or painted: large buy orders placed and then cancelled are the classic fake wall.
- New dragon-tiger seat detail — the branch seats behind a stock appearance (up to 10), with buy / sell / net amounts and a hot-money flag. More specific than an institution summary: you see whether it was Stock Connect, a named brokerage branch, or hot money.
- Use them together: the queue shows orders that may be cancelled, while large-order fills show what actually traded.
- The queue covers the current day only; dragon-tiger data is post-close, so a blank
datereturns the most recent appearance (check the returneddate). - The queue is pending-order data, not fills; seat detail returns at most 10 rows.
- Who can use it: Unlimited (¥199/mo) only.
2026-10-18EndpointsNew: L2 large-order fills with active/passive split (Unlimited tier)- New L2 large-order fills — every fill above a value threshold for a single stock today, each tagged with one of four directions: active buy / passive buy / active sell / passive sell.
- Parameters:
money0-4 maps to 300k / 500k / 1M / 3M / 10M CNY thresholds;limitup to 200 fills. - What it answers: whether someone is aggressively accumulating (active buys dominate), absorbing offers (passive buys), or distributing (concentrated active sells) — a net figure alone cannot tell these apart.
- The endpoint covers the current day only.
totalis the number of qualifying fills;truncated=truemeans the list was cut bylimit, sosummarycovers only the returned sample. - This is fill data (executed trades), not order-by-order data and not a 10-level order book.
- Who can use it: Unlimited (¥199/mo) only.
2026-10-18EndpointsNew: L2 large-order flow (Unlimited tier)- New L2 large-order flow — a per-minute net large-order curve for a single stock on the current day (
time+net, in CNY); positive means net buying, negative net selling. - What it answers: whether large orders are steadily accumulating or withdrawing, which minute saw concentrated selling, and the intraday rhythm of large-order money.
- The endpoint covers the current day only — on non-trading days or before the open it returns the most recent trading day in full; use the
dayfield. During the session the point count keeps growing (about 242 points for 09:30–15:00). - This is a large-order aggregation (by trade value), not order-by-order data and not a 10-level order book.
- Who can use it: Unlimited (¥199/mo) only; other tiers get a clear message with an upgrade link.
2026-10-18EndpointsFive new MCP tools — company profile / industry chain / event calendar / margin / block trades- What changed: the MCP tool list now includes
ashare_profile(company profile),ashare_industry_chain(183 industry-chain themes),ashare_calendar(corporate event calendar),ashare_margin_trade(margin trading detail) andashare_block_trade(block trade detail) — 49 → 54 tools. - All five capabilities are now callable directly from MCP clients — previously you had to call the REST endpoints yourself; now the AI can use them as soon as the server is configured.
ashare_profile: listing date / main business / industry / website / registered capital / chairman — use it for "what does this company do", andashare_snapshotfor "is it expensive". Cached for 24 hours.ashare_industry_chain: 183 themes;mode=listfor themes,graphfor the chain graph (upstream / midstream / downstream),stockfor a stock position (with relevance and business description).ashare_calendar: dividends / lock-up release / earnings dates / IPOs / meetings / placements over roughly one month. The response is multi-section, and a section is omitted entirely when it has zero events that day (not lost data). This is not a macro calendar.ashare_margin_trade: financing balance / buy / repay plus short-selling detail;codeaccepts a comma list. Margin data is disclosed T+1, so today is not available today.ashare_block_trade: trade price / premium-discount rate / buyer and seller branches; leavecodeempty for the whole market. When there are no block trades that day the response falls back to quote metadata (proving the symbol is fine) — check for trade fields to decide whether any block trade happened.- Who can use it: these five are not Pro-only — they work from Trial (¥9.9) upward.
2026-10-18EndpointsNew batch technical-indicators endpoint (Pro) — 55 indicators for a whole pool in one call- What changed: new
/v1/technical-batch— MA / EMA / MACD / KDJ / RSI / BOLL / BIAS / WR / DMI / OBV / VR / BBI / TRIX / DPO / PSY / ENE / CCI and more, 55 fields in total, for up to 500 symbols per call (1000 on Unlimited), with per-symbol status. Each item has exactly the same shape as a single row of/v1/technical. - How it worked before:
/v1/technicaltakes one symbol at a time, so scanning 500 symbols meant 500 calls (about 8 minutes, plus rate-limit backoff); this returns them in one call — 500 symbols in 1.2 seconds, 1000 in 1.8 seconds (measured). - Why not just use
/v1/technical: itscodeaccepts commas too, but the response is one wide table with no per-symbol status, and invalid symbols silently vanish, so you cannot tell which one is missing. This endpoint fills in every requested symbol and says why any are missing. - Billing: per symbol actually returned; failures are not charged.
- This is the only batch endpoint without a data-source bottleneck: measured 3000 symbols in 2.8 seconds, 100% returned, zero rate limiting, so the cap is set by response size (55 columns x N rows) — hence 500 / 1000, wider than the 200 / 300 of the other batch endpoints.
- Widest market coverage: A-shares (including Beijing exchange), HK and US symbols can be mixed in one call, and indices work too (
sh000001) — the other batch endpoints cannot do this. - ⚠️ Daily bars only: the period parameter has no effect (week / month / m30 return exactly the same numbers as daily), so this endpoint does not expose a period parameter — otherwise you might think you are getting weekly MA5 when it is actually daily.
- History:
date=YYYY-MM-DDreturns the indicators for that trading day (applies to all symbols at once); non-trading days fall back to the nearest session — trust thedatefield in the returned rows. - Who can use it: Pro (¥99) and above.
2026-10-18EndpointsNew batch financials endpoint (Pro) — all three statements for a whole pool in one call- What changed: new
/v1/finance-batch— income statement / balance sheet / cash flow for up to 10 symbols per call (20 on Unlimited), withperiods(default 4),type(one statement only) andfields(all186 columns /core69), plus per-symbol status. - How it worked before:
/v1/financetakes one symbol at a time, so screening 10 symbols meant 10 calls; this returns them in one call — 10 symbols x 4 periods in 4.2 seconds (measured). - Why not just use
/v1/finance: itscodeaccepts commas too, but the response is three flat tables with no per-symbol status — you cannot tell which symbol is missing or why. The new endpoint fills in every requested symbol and says why any are missing. - How billing works: charged per symbol actually returned — failures are not counted, and rate-limited calls are refunded automatically.
- ⚠️ Limits: 10 symbols on Pro / 20 on Unlimited, and symbols x periods must be <= 40: quota is consumed per symbol-period - 10 symbols x 4 periods was complete three times, 20 x 2 was complete; but 20 x 4 returned only 5, and 50 was rate-limited outright. For more, split into batches (a few seconds apart).
- ⚠️ A-shares only: HK / US symbols are unverified in batch mode, so they are flagged as failed with a pointer to the single-symbol endpoint.
- Who can use it: Pro (¥99) and above.
2026-10-18EndpointsNew batch shareholder-research endpoint (Pro) — see who holds a whole pool in one call- What changed: new
/v1/shareholder-batch— top-10 holders / top-10 float holders / holder count for up to 200 symbols per call (300 on Unlimited), with per-symbol status; each entry has exactly the same shape as/v1/shareholder. - How it worked before:
/v1/shareholdertakes one symbol at a time, so scanning 200 symbols meant 200 calls (about 3 minutes, plus rate-limit backoff); this returns them in one call — 200 symbols in 4.1 seconds / 300 in 5.6 seconds (measured). - Why not just use
/v1/shareholder: itscodeaccepts commas too, but the response is one big block of markdown with no per-symbol status, and invalid symbols silently vanish — you cannot work out which one is missing. The new endpoint fills in every requested symbol and says why any are missing. - How billing works: charged per symbol actually returned — failures are not counted.
- ⚠️ A-shares and HK: Beijing exchange / ETF / index / US symbols return nothing in batch mode, so they are flagged as failed with a pointer to the single-symbol endpoint.
- ⚠️ A-shares and HK differ in section meaning: A-shares give top-10 holders / top-10 float holders / holder count, HK gives shareholders / holder distribution / institutional holdings — read them by
marketandslug, never by position. - Who can use it: Pro (¥99) and above.
2026-10-18EndpointsNew batch chip-distribution endpoint (Pro) — scan a whole pool for trapped positions and cost- What changed: new
/v1/chip-batch— profit ratio / average cost / 90% and 70% concentration for up to 200 symbols per call (300 on Unlimited), with per-symbol status;datelets you look back over roughly the last year. - How it worked before:
/v1/chiptakes one symbol at a time, so scanning 200 symbols meant 200 calls; this returns them in one call — 200 symbols in 3.9 seconds (measured). - Why not just use
/v1/chip: itscodeaccepts commas too, but the response is a wide diagonal matrix (row i only carries symbol i, everything else is-), and the whole column group of an invalid symbol disappears — you cannot work out which one is missing. The new endpoint fills in every requested symbol and says why any are missing. - How billing works: charged per symbol actually returned — failures are not counted.
- ⚠️ A-shares only: HK / US symbols and sector codes are unreliable in batch mode, so they are flagged as failed with a pointer to the single-symbol endpoint.
- Who can use it: Pro (¥99) and above.
2026-10-18EndpointsNew batch company-profile endpoint (Pro) — see every company at once after screening- What changed: new
/v1/profile-batch— listing date / main business / industry / website / registered capital / chairman / contact details for up to 100 symbols per call (150 on Unlimited), with per-symbol status. - How it worked before:
/v1/profiletakes one symbol at a time, so screening 100 symbols meant 100 calls; this returns them in one call — 100 symbols in 6.8 seconds (measured). - Why not just use
/v1/profile: itscodeaccepts commas too, but the rows come back mixed together with no per-symbol status, and invalid codes are silently dropped — you cannot tell which one is missing. The new endpoint fills in every requested symbol and says why any are missing. - How billing works: charged per symbol actually returned — failures are not counted.
- ⚠️ A-shares only (including indices and ETFs): HK / US symbols and sector codes are unreliable in batch mode, so they are flagged as failed with a pointer to the single-symbol endpoint (never wrong data).
- Who can use it: Pro (¥99) and above.
2026-10-18EndpointsFour new batch endpoints (Pro) — stop writing loops to scan a watchlist- What changed: three new batch endpoints —
/v1/quote-batch(quotes),/v1/kline-batch(K-lines) and/v1/snapshot-batch(full profile: valuation / market cap / share capital and more). The existing/v1/fund-batch(money flow) is now also exposed through MCP clients. - How it worked before: fetching 300 symbols meant 300 separate calls — about 0.7s each, 3.5 minutes in total, and rate limiting could interrupt it midway. Now one call returns them in around 4 seconds.
- What you get back: one record per symbol, each flagged as ok or failed (ask for 300, get 300 entries), plus
requested/returned/failedcounts. - How billing works: charged per symbol actually returned — failures and suspended stocks are not counted.
- Limits: Pro up to 200 symbols per call; Unlimited 250–1000 depending on the endpoint.
- Who can use it: Pro (¥99) and above. Standard does not include batch endpoints.
- ⚠️ Intraday (
/v1/minute) does not support batching — the data source only accepts one symbol per call.
2026-10-17FixAnonymous multi-symbol limit now returns 422 instead of 401- What changed: when an anonymous call (no API key) passes more than 5 symbols to
/v1/klineor/v1/quote, the status code is now422instead of401. The response body is unchanged (max_codes/requested_codes/pricing). - Why:
401means "unauthenticated", which clients often read as an invalid key and retry automatically. The actual cause is too many symbols in one request (a parameter problem), so422is the accurate code. - Who is affected: anonymous batch calls only. Calls with an API key, and anonymous calls with 5 symbols or fewer, are completely unaffected.
- Action needed: if your code treats
401as "need to buy an API key", switch that check to422(or read theerrorfield in the response body). - 📌 For contrast: exhausting the daily anonymous item quota still returns
429(with aRetry-Afterheader) — a different kind of limit, handle them separately.
2026-10-17Endpoints/v1/chip gains date for historical chip data · /v1/etf gains kind (detail / holdings / nav / company / holders / financial)/v1/chipnow acceptsdate(YYYY-MM-DD, defaults to the latest session): look up chip metrics for any trading day (average cost / 70% and 90% concentration / profit ratio / closing price).- ⚠️ Chip history is retained for roughly the last year only (verified: data starts 2025-10-24). Earlier dates return an empty result — the response is
ok:falseand no quota is consumed. - ⚠️ Non-trading days and future dates fall back to the most recent trading day (e.g.
date=2026-10-17, a Saturday, returns 2026-10-16 data) — always read thedatefield of the returned row instead of assuming it equals the date you passed. /v1/etfnow acceptskind:概览(default, identical behaviour to before) ·详情(fund facts / fees / shares) ·持仓(constituents and weights) ·净值(historical NAV, requires bothstartandend; returnsnavwith 4 decimals plusclosePrice, so you can derive the premium/discount yourself) ·公司(manager / custodian / tracked index) ·持有人(holder accounts / institutional vs retail split) ·财务(stock / bond / commodity / fund allocation).- ⭐
kind=净值returns the real unit NAV (4 decimals), not the closing price — for instruments with a large premium/discount (cross-border or commodity ETFs) the two can differ by more than 10%, so do not mix them up. - ⭐ Codes may be passed bare (
510300) or prefixed (sh510300); both forms return identical results. - ⭐ Read fields from
structured/tables:dataremains the formatted text from the data source (unchanged),structuredis the first table andtablesholds every table. Forkind=详情(3 tables) andkind=公司(2 tables) onlytablesis complete (it includes the fund-manager section). - ⚠️
/v1/etfaccepts ETF codes only: passing an equity code now returnsok:falseand consumes no quota (it no longer returns an all-zero shell) — resolve the code first with/v1/searchor/v1/hot?type=etf. - Do I need to change my code? No. Without the new parameters, both endpoints respond byte-for-byte identically to before.
- ⚙️ The
ashare_etfMCP tool now also acceptskind/start/end(tool count unchanged at 39).
2026-10-16EndpointsAnonymous batch requests: /v1/kline and /v1/quote now allow at most 5 symbols per call- What changed:
codeon/v1/klineand/v1/quoteaccepts a comma-separated list; from now on anonymous calls (no API key) can pass at most 5 symbols per request, otherwise the call returns401(the response includesmax_codes: 5). ⚠️ That status code was corrected to422on 2026-10-17 — see the newest entry above. - Who is affected: anonymous calls only. Calls with an API key are unaffected — the symbol limit is unchanged for them.
- Do I need to change my code? No. Querying a single symbol (the most common case) works exactly as before.
- Also added:
/v1/klinenow enforces a per-request row limit — requests estimated above 60,000 rows return413. With typical parameters (e.g.count=30) and a single symbol this is not triggered. - Need batch access? An API key removes the symbol limit: https://ashareapi.com/pricing
2026-10-16FixCorrected the stage field of /v1/auction — its meaning is not defined; do not rely on it- ⚠️ Documentation correction: the
stagefield returned by/v1/auctionwas previously described as "auction phase (0 = cancellable 09:15–09:20 · 1 = non-cancellable 09:20–09:25)" (see the 2026-10-15 entry). That description does not hold. Measured (2026-10-16, 10 symbols, 550 rows): 35% of rows in 09:15–09:19 (cancellable) are1; 59% of rows in 09:20–09:24 (non-cancellable) are0; and at 09:24 (which should always be1) 7 out of 10 symbols return0. - ⭐ Correct usage: this field is a source-side marker, and the source's published documentation does not define what its values mean — so do not rely on it. To split the cancellable / non-cancellable windows, use
time: 09:15–09:20 cancellable · 09:20–09:25 non-cancellable (a fixed market-wide rule). - ✅ Unaffected:
time/price/volume, plusis_limit_up/up_px/down_px/open_px/preclose_px/hprice. The field is still returned, so existing code keeps working.
2026-10-15EndpointsNew /v1/auction — call auction (intraday series / historical ticks / volume + "touched limit-up" flag)- New endpoint
/v1/auction: the 09:15–09:25 call auction for a single stock, including a "touched limit-up during auction" flag. modehas 3 options:stockintraday auction series (default, minute ticks covering 09:15–09:24) ·stock-historyhistorical auction series (second-level ticks, requiresdate) ·volumeintraday auction volume / net amount (single point).- ⭐
is_limit_up= the "touched limit-up" flag — true when the auction high price reaches the limit-up price. The limit-up price is taken from a raw field of the data source (not computed by us), so the differences between main board 10% / ChiNext·STAR 20% / ST 5% / Beijing Exchange 30% are correct by construction — no need to apply your own formula. - Response: each row has
time/price/stageauction phase (0 = cancellable 09:15–09:20 · 1 = non-cancellable 09:20–09:25) /volumecumulative volume; top-level fields includeup_pxlimit-up price /down_pxlimit-down price /open_pxthe 09:25 match price /preclose_pxprevious close /hpriceauction high. - ⚠️
mode=stockcovers today only (data-source limitation); for history usemode=stock-historywith adate(must be a real trading day). - ⚠️ Tier: Standard (¥29.9) and above.
- ⚙️ MCP now exposes
ashare_auction(tool count 38 → 39).
2026-10-14FixTier gate added for intraday and minute bars (Standard and above) · real-time quotes for paid tiers- ⚠️ Breaking change:
/v1/minute(intraday trend) and the minute bars of/v1/kline(period=m1–m120) now require the Standard tier (¥29.9) and above;trial(¥9.9) no longer has access. - ⭐ Why: both are intraday real-time data, and the tier gate had been left unregistered — so the trial tier could call them, breaking the site-wide convention that time-sensitive endpoints start at Standard (the two entry points of the same capability shipped together were tightened together).
- ✅ Real-time quotes for paid tiers:
/v1/quote,/v1/klineand/v1/minuteare now uncached for paid users (every call returns the latest value); free / anonymous keeps a 5-minute cache to protect service stability. - Unaffected: the daily / weekly / monthly / quarterly / yearly bars of
/v1/klineremain free (no key required); fields and semantics are unchanged.
2026-10-14DataTiered quote freshness — paid tiers are uncached, every call returns the latest value- ⭐ Paid tiers are uncached:
/v1/quote(real-time quote) and the daily / weekly / monthly bars of/v1/klinereturn the latest value on every call — repeated intraday calls reflect the moving price. - ⭐ Standard (¥29.9) and above get the intraday data uncached too: the minute bars of
/v1/kline(period=m1–m120) and/v1/minute(intraday trend). - Free / anonymous keeps a 5-minute cache: to protect service stability (free traffic is the majority and mostly polling), free-tier quotes refresh on a 5-minute cycle.
- ⭐ The data itself has always been current-day (including the intraday bar for today); fields and semantics are unchanged.
- ⚠️ Note: quotes are rolled during the trading session; historical daily bars (already-closed sessions) are unaffected.
2026-10-14EndpointsNew /v1/review and /v1/limit-up — short-term review / limit-up pool (structure + detail)- 2 new endpoints:
reviewfor structure,limit-upfor detail — two complementary views covering a short-term trader's daily needs. /v1/review(review view),modehas 2 values:boardsector mainlines (where the money went today: sectors +limit_countlimit-ups +amountsector turnover +stocksconstituents) ·ladderconsecutive-board ladder (grouped by board height, withboard_name/market_cap/tipse.g. "7天4板" /max_streak/limit_days/is_repackreverse flag)./v1/limit-up(limit-up view),modehas 1 value:limitlimit-up pool detail — the full set of limit-up stocks in one call (all board heights merged), each withreasonlimit-up reason /seal_amount/seal_amount_max/main_net/main_buy·main_sell/amount/concepts/float_cap/turnover/streak/pct.- ⭐ Both accept
date(YYYY-MM-DD) for a historical trading day; empty = realtime. History must be a real trading day (otherwise empty). - ⭐
datais always a row array (DataFrame-ready); the daily total forlimitis the top-levelcount. - ⚠️ Tier: Standard (¥29.9) and above.
- ⚙️ MCP gains the
ashare_review·ashare_limit_uptools (total tools 36 → 38).
2026-10-14ReferenceNew “AI stock screening” landing page — factor → screen → backtest in three steps- New /en/ai-screener/: an AI stock screening landing page — how to screen with our API: 88 factors → expression screening → 29 backtested strategies, the whole chain from one API.
- ⭐ Each step maps to an endpoint: /v1/factors to define factors · /v1/screen to write conditions · /v1/backtest to validate (with real trading costs). Comes with a runnable screening expression.
- ⚠️ Boundaries up front: we do not pick stocks · do not promise returns · do not touch automated trading — what you get is the screening capability (reproducible), not the picks.
- Sibling page: Agent quant (category definition · checkable data).
- Available in both languages: 中文页.
2026-10-13EndpointsNew /v1/abnormal — abnormal moves / deviation (short-term regulatory signals)- New endpoint
/v1/abnormal: short-term regulatory abnormal moves / deviation — stocks triggering a *severe move*, a *10-day 100% deviation*, or on the abnormal-move watchlist. These are key short-term risk signals (a stock flagged for large consecutive gains or drops). modehas 7 values:deviationdeviation hot list (today, default) ·deviation_hist(history, needsdate) ·deviation_plan(both today and tomorrow, split via thegroupfield) ·abnormalabnormal-move frequency + reason (needsdate, full daily set) ·watchwatchlist (today) ·watch_hist(history, needsdate) ·alertsevere-move alerts (today,reasonis a human-readable trigger).- ⭐ Fields:
reasontrigger-rule description (e.g.10日100%) ·pctchange % ·turnoverturnover % (deviation lists) ·board_descconsecutive-limit description ·conceptsconcepts ·windowdeviation window ·rule2second rule ·trigger_pcttrigger threshold % ·limit_pricereference price ·statusstate ·datedata date ·pricelast price ·abn_startabnormal-move start date (0000-00-00when not triggered). - ⭐
datais always a row array (DataFrame-ready); the daily total forabnormalis the top-leveltotal, and eachdeviation_planrow carriesgroup(today/tomorrow). - ⚠️ For history modes,
datemust be a real trading day (otherwise it returns empty); tier: Standard (¥29.9) and above. - ⚙️ MCP gains the
ashare_abnormaltool (total tools 35 → 36).
2026-10-13FixFixed /v1/search returning empty for most A-share names / codes- Issue:
/v1/searchreturned an empty result for most A-share names (e.g. Mao Tai, Wu Liang Ye, BYD) and codes (e.g.600519) — yet confirming a code from a name is exactly what this endpoint is for, so its most common use was broken. - ⭐ Root cause: the underlying search command was given combined category switches (stock + fund + sector), whose real semantics are an intersection, not a union — so any keyword not matching all three categories came back empty.
- Fix: search no longer restricts by category — it now returns every match at once (stocks / funds / sectors / Hong Kong / US).
- ✅ Verified (paid tiers):
Mao Tai→sh600519·Wu Liang Ye→sz000858·Tai Ji Shi Ye/600667→sh600667·Ping An→ Ping An Insurance + Ping An Bank + HK/US listings — all returned correctly.
2026-10-13GuidesDeveloper page now covers QMT / Ptrade integration (copy-paste examples)- New "Connecting to QMT / Ptrade" section on the developer page: both terminals ship an older built-in Python (QMT 3.6 · Ptrade 3.5) with broker library whitelists and sandboxes, so third-party packages generally will not install.
- ⭐ The workable pattern: fetch outside, read files inside — pull data with the SDK outside the terminal into CSV, then let the strategy read local files with the terminal’s bundled pandas (no installs, no network). Two copy-paste examples included.
- ⚙️ Also: the Python / Node install commands on the docs and developer pages are now a tab switcher (previously shown side by side).
- New /en/developers/ page: copy-paste starters for Python and Node.js, a table of common methods, and "what you can build with it" (local research / dashboards / strategy scripts / batch scanning).
- ⭐ New "Developers" item in the navigation, plus SDK sections on the home and pricing pages — previously the SDK only appeared inside a feature list, with no entry point of its own.
- ⚙️ Docs updated: the quick start on /en/docs/ now covers both curl and the SDK (previously curl only).
- Official SDKs:
pip install ashareapi·npm install ashareapi— free endpoints work right after install, no key needed.
2026-10-13GuidesDocs "Quick start" now offers both curl and the official SDKs — two lines to a DataFrame- New official SDK section on the docs page: Python (
pip install ashareapi) and Node.js / TypeScript (npm install ashareapi), each with its install command and a two-line example, plus one-click copy. - ⚙️ Onboarding now covers both paths: the docs page previously only showed curl. Now curl and the SDKs sit side by side — keep using curl to call endpoints directly; use an SDK for scripts, dashboards or strategy code. The response shape is identical.
- ⭐ Both packages are published on PyPI and npm; the package pages are the source of truth for versions and usage.
2026-10-12EndpointsNew quant-only batch endpoint /v1/fund-batch — fetch capital flow for a whole batch of stocks in one call (Pro / Unlimited)- New /v1/fund-batch: a batch capital-flow endpoint — one call returns capital flow (main money flow + dragon-tiger list + block trades + margin) for many stocks, comma-separated as
codes=sh600519,sz000001,sz300750. Built for people doing quant who need to scan a pool, instead of calling/v1/fundone symbol at a time. - ⭐ Per-call limit: Pro 200 symbols (about one sector) · Unlimited 1,000. Billed per symbol actually returned — failures are not charged.
- Per-symbol results: the response carries
requested/returned/failedcounts and anokflag per symbol — so you can see exactly which ones came back (capital flow covers A-shares only; Hong Kong has main net flow only, US has none). - A separate endpoint: fully independent of the single-symbol
/v1/fund— existing endpoints are untouched. - Also in Chinese: 中文更新日志.
2026-10-12ReferenceNew “Agent quant” category page — how it differs from traditional quant, and why the data must be checkable- New /en/agent-quant/: a category definition page for Agent quant — what it is, how it differs from traditional quant (JoinQuant / RiceQuant / Uqer / MyQuant), and why “did the agent make it up” is the real bottleneck of the category.
- ⭐ The core claim: in traditional quant a bug can be debugged; in Agent quant you cannot tell computed from invented ⇒ so what it needs is not more data but checkable data (fixed forward adjustment / consistent units / empty result ≠ failure / fixed envelope).
- ⚠️ Boundaries first: no returns promised · no automated trading · no competing on “how accurate the AI’s picks are”.
- How to do it lives in /en/docs/ai-quant/ (four data paths compared · REST and MCP · the four traps that fool a backtest); this page only defines the category.
2026-10-12GuidesPer-platform pages arrive for the China platform guide — Doubao first, the rest to follow- New /en/docs/mcp-china/doubao/: a dedicated Doubao setup page — install steps, key placement, example questions and FAQ in one place, easier to follow than the one-line summary on the overview.
- Every platform card on the overview /en/docs/mcp-china/ now links to a “Full guide →” page.
- ⭐ One page per platform: Bailian / Coze / Yuanqi / Tencent Cloud / Volcengine Ark follow the same template at
/en/docs/mcp-china/<platform>/.
2026-10-12GuidesChina platform guide now covers Doubao Work — custom headers supported, so paid tools work too- /en/docs/mcp-china/ now includes steps for Doubao Work: left sidebar Plugins · Skills · Partners → + Add on the Plugins page → New custom plugin → transport HTTP, server URL
https://api.ashareapi.com/mcp→ add a Custom Header (nameAuthorization, valueBearer <your key>) → save, then type/or@in a task to search forashareapi. - ⭐ Doubao supports custom request headers (its official "Using plugins in Doubao Work" doc documents HTTP + custom headers) ⇒ paid tools work too — unlike Bailian / Coze, where only free tools are available.
- ⚠️ Doubao calls this a "plugin" (entry: Plugins · Skills · Partners); "connector" in older third-party guides is the same feature under an old name. If it first reports a failure, open the plugin list and click Restart.
- Also in Chinese: 中文更新日志。
2026-10-12ReferenceQuick start and pricing pages now link a "code format" reference — codes need a lowercase market prefix/docs/and/pricing/each gained a "code format" entry linking to the code format reference:sh600519(Shanghai) ·sz000001(Shenzhen) ·bj920002(Beijing) ·hk00700(Hong Kong) ·usAAPL(US).- ⭐ Why: quote endpoints accept only "lowercase prefix + code". Measured on 2026-10-12 — the five correct forms above all return data, while
600519(no prefix),600519.SH(dot notation),SH600519(uppercase prefix) andsh60051(wrong length) all return empty (ok:false+ empty array). It looks like "no such ticker", but it is a format problem. - Always include the prefix; if unsure, look the code up by name with
/v1/search. - Also in Chinese: 中文更新日志。
2026-10-12DataPlaybook library /v1/playbooks expanded: 13 → 25 playbooks (new execution & risk-control entries)/v1/playbooksgrew from 13 to 25 playbooks, adding trade execution and risk-control methods: dip-buy screening · pullback opportunity · platform breakout candidates · breakout execution · failed-breakout exit · key-level alerts · target-reached handling · staged take-profit · stop-loss discipline (price/time/thesis) · position sizing · trim-or-hold · adding to a winner.- Every entry keeps the same shape:
signals(formula + data + threshold),stages,actionfor each stage,avoids(what losses it helps you dodge) andevidence(backtest figures; entries not yet run are marked "pending" — we do not invent numbers). - Data boundaries stated honestly: some sentiment-cycle signals need limit-up / failed-limit pools, which we currently cover only as *counts* (
/v1/changedist) — thedata_needsfield says what substitutes and what is missing. - Themes unchanged (still 12): new entries slot into the existing risk control and technical timing categories.
- Paid tier — Standard (¥29.9) and above; reading the list does not consume backtest quota.
2026-10-11Data/v1/notice (filings) lowered to Standard (¥29.9) — the news capability is now complete at Standard/v1/notice(company filings) moves from Pro to Standard: now available from Standard (¥29.9) and up, reaching about 16 years back (Moutai: 929 filings to 2010 in our tests).- ⭐ Why: Standard previously had only
news(latest) and was missing the "history" half — putting news + filings in the same tier means upgrading to Standard gives you a complete news picture, without having to buy another tier above it. - Pro (¥99) keeps its research-grade capabilities unchanged: batch backtest (many stocks × many strategies side by side) and full-market single-day K-line (5,000+ symbols in one call) — these remain what Pro can do that lower tiers cannot.
- Existing Standard customers get this automatically, no action needed; calls and billing are unchanged (filings count toward usage like any other data endpoint).
2026-10-11DataStrategy library now carries measured performance — real results for all 29 strategies, losers included/v1/strategiesgains aperformanceblock per strategy: its real result across 800 stocks (CSI 300 + CSI 500) since 2019 — mean/median return, Sharpe, max drawdown, share of profitable symbols, and trade count. Portfolio strategies (9) are backtested on the full market (all A-shares with available data).verdictgrading:effective/neutral/weak/avoid(measured loss) /baseline(reference) — so you can spot the good and the bad at a glance.- ⭐ Losing strategies are listed too: ⚠️ our testing shows most classic technical-indicator strategies do not hold up once real costs are included (e.g. Williams %R is profitable on only 20% of symbols). We show this plainly so you can avoid the common traps instead of only seeing the flattering ones.
noteexplains the pitfall — the mechanism behind each failure (e.g. "MA cross whipsaws in range-bound markets ⇒ 1,454 trades over 8 years, costs eat the edge") rather than a vague good/bad label.- ⭐ Portfolio strategies carry a "period consistency" check (
consistent+by_period): the returns across four windows (full / first half / last 4y / last 2y) side by side, so you can see whether the direction holds. ⚠️ Our testing found strategies likesmallcapflip sign when the window changes (first half -40% / last 4y +175%) — direct evidence of backtest overfitting — so they are automatically downgraded toneutralinstead of being presented to you as "effective". - Caliber stated explicitly (
cost_included: true): commission (0.03%), stamp duty (0.05%), slippage (0.1%) are all included; the sample includes delisted stocks (no survivorship bias); the period starts 2019-01-01. - Tier: Standard (¥29.9) and above; listing does not consume backtest quota.
- ⚠️ Historical results do not indicate future performance; research use only, not investment advice.
- Also available in Chinese: 中文更新日志。
2026-10-11EndpointsTop-trader boards expanded to 6 types — added market-wide board / high-win-rate buys / high-win-rate seats/v1/lhbnow accepts 6typevalues: addedall(market-wide board — every listed name that day, with net buy / buy / sell amounts),winbuy(high-win-rate buys — net amount / win rate / seat detail) andwinseat(high-win-rate seats — seat win rate + linked symbols). The existinginstitution(institutions) andactiveseat(active seats) are unchanged, so current calls keep working.- New
dateparameter:/v1/lhb?type=all&date=2026-10-08fetches a specific trading day; omit it for the latest trading day. hotmoney(hot-money board) has no data available — it now returns guidance pointing you toall/activeseat/winbuyinstead of an empty result, so you can switch quickly.- Availability: any paid tier (from Trial ¥9.9; anonymous requests without a key are not allowed). Priced the same as the existing top-trader board call.
- Also available in Chinese: 中文更新日志.
2026-10-10EndpointsNews coverage goes live: /v1/news (latest) · /v1/notice (filings, ~20 years back) · /v1/hot now serves 5 leaderboards- New news for a symbol:
/v1/news?code=sh600519returns a list of financial news (title + source + time);/v1/news?id=SN...returns the full body of that article — so an AI can actually read the story. Use it when a user asks "what is the news on this stock" or "what does this article say". - New filings / announcements:
/v1/notice?code=sh600519returns the announcement list, supports type filtering and offset paging, and reaches about 20 years back (a large-cap name pages all the way to 2006); passidto fetch one announcement. The detail has two shapes: most short filings return the body intext; a few long reports (100+ page interim/annual filings) return a PDF link inurl— fetch and parse it yourself (for structured figures use/v1/finance). /v1/hotnow serves 5 leaderboards via the newtype=parameter. Defaults tostock, so existing calls are unaffected.- ⚠️ The five types have different field shapes — read the fields that match your
type:stocktrending stocks (code/name/zdfchange/zxjlast) /wxwhole-market heat (same, plus apvheat value) /newstrending articles (news_id/news_title/source, ⛔ nocode) /sectorhot sectors (symbol/rankdelta, ⛔ nocode) /etftrending ETFs (title,tag). - Freshness vs. history:
/v1/newscovers the latest items (the data source offers no paging). For historical depth (filings and disclosures) use/v1/notice. - Tiers:
/v1/hotis free (no key required);/v1/news(latest news) starts at Standard ¥29.9;/v1/notice(historical filings) starts at Pro ¥99. - How to query:
/v1/news?code=sh600519returns the list → pass itsidto/v1/news?id=...for the body (two calls, because full text is fetched per article). - Bodies are AI-friendly: news bodies are cleaned into plain Markdown (HTML tags, image placeholders and dead internal links removed, stock-code links kept as
茅台(sh600519)), ready to feed straight to an LLM — no markup noise, fewer tokens. Timestamps are normalised toYYYY-MM-DD HH:MM:SS. - New MCP tools
ashare_newsandashare_notice;ashare_hotgains atypeparameter (tool count 33 → 35).
2026-10-10EndpointsNew whole-market endpoint /v1/market-kline — every A-share's daily bar for one day in a single call- New whole-market single-day K-line:
/v1/market-kline?date=2026-09-30returns every A-share (5,000+ symbols) for one trading day in a single call, replacing the old "loop over 5,000 symbols" pattern — call it once after the close to update the whole market. - Built for: local market databases, whole-market scans and daily batch updates. Requests drop from 5,000+ to 1; each response is about 785 KB.
- Available on the Pro (¥99) and Unlimited (¥199) tiers: Standard and below cannot use it. A separate daily cap applies (Pro 200/day · Unlimited 1,000/day).
- Billing: because one request moves the data of 5,000+ symbols, one call counts as 500 calls of quota — still far below the 5,000+ calls a symbol-by-symbol loop would cost.
- Scope: A-share stocks only (5,572), from 2016-08-08, raw unadjusted prices; ETFs, convertible bonds, indices, sectors, Hong Kong, US, futures and FX are not included. Data is maintained daily from the official feed, so the latest available day is the previous trading day (intraday same-day data is not available).
- New MCP tool
ashare_market_kline(tool count 32 → 33).
2026-10-10EndpointsNew doc-search endpoint /v1/wiki — agents can verify field units and conventions instead of guessing- New doc search:
/v1/wikiserves the product documentation — field units, conventions, common misuse and cross-market differences. When an AI Agent is unsure what a field means or which unit it uses, it can look it up rather than guess. - Progressive loading to save context: no params returns the catalog (one line per doc);
q=keywordreturns matching docs as summaries (with a snippet, not the full text);slug=xxxreturns that one doc in full. It never dumps every doc at once — go catalog first, then fetch the single doc you need. - New MCP tool
ashare_wiki_search(tool count 31 → 32). It is a free tool (no key required) — same free tier as quote / kline / hot / market overview / breadth.
2026-10-08EndpointsQuant upgrade: batch backtest /v1/backtest-batch · portfolio weighting · IP allow-list /v1/ip-whitelist- New batch backtest:
/v1/backtest-batchruns many stocks × many strategies in one call and returns a side-by-side comparison table, so you no longer loop over single calls. Up to 20 stocks × 10 strategies, max 60 combinations; each stock's history is loaded once and shared across every strategy, so 15 combinations come back in about a second. One bad stock does not abort the batch (that row records the reason, the rest still run). Paid tier (Pro and above). - Portfolio backtest now supports a weighting scheme:
/v1/backtest?mode=portfolioacceptsweighting=inv_vol— position size is the inverse of each stock's volatility (steadier names get more, wilder names get less), alongside the default equal-weightequal. On the same portfolio and window, inverse-volatility weighting lost 11.85 percentage points less than equal weight (and drew down less too). Unlimited tier. - New IP allow-list: lock an API key to a fixed set of IPs, so a leaked key cannot be used from anywhere else.
POSTsets it (up to 50) ·GETreads it ·DELETEclears it. Once set, calls from other IPs get403; your current IP always passes, so you cannot lock yourself out. Unlimited tier. - Unlimited tier gains a dedicated concurrency slot: heavy requests (backtests, full-history K-line) get their own slot instead of competing with other tiers — while others queue at peak, Unlimited keeps running. The most direct difference from Pro.
- Backtest daily quota now scales by tier: Standard 30/day · Pro 300/day · Unlimited 1,000/day — backtests are a CPU-heavy endpoint, so the quota scales with the tier.
- Portfolio backtest supports several ranges: no longer a single fixed window — there are four preset ranges (full period / last 4y / last 2y / first half). Endpoints follow the latest trading day, so they never fall behind; the one fixed end date is a deliberate out-of-sample control. An unsupported range returns the list of available ranges, so you do not have to guess.
- Batch backtest is quota-limited per tier: Pro 20 batches/day · Unlimited 100/day; up to 20 stocks × 10 strategies per call.
- Endpoint count 38 → 40.
- Both languages available: see the English changelog.
2026-10-08EndpointsMCP tools now cover quant (5 new tools · 31 total)- Five new MCP tools:
ashare_backtest(quant backtest),ashare_backtest_batch(batch backtest),ashare_strategies(strategy list),ashare_factors(factor library) andashare_playbooks(playbook library) — so in Claude Code, Cursor and other clients you can just say "backtest this strategy for me" and the AI calls them itself. Tool total 26 → 31. - Tiers: strategy / factor / playbook lists and single-stock backtests are Standard (¥29.9) and above; portfolio and batch backtests are Pro (¥69) and above. Check your remaining quota with
ashare_usage. - Also in Chinese: 中文更新日志.
2026-10-08EndpointsNew quant backtest /v1/backtest · quant strategy list /v1/strategies · quant factor library /v1/factors · quant playbook library /v1/playbooks- New quant backtest:
/v1/backtestruns a built-in strategy over history and returns total return / CAGR / Sharpe / max drawdown / annualised volatility, optionally with the daily equity curve. Costs are included (commission, stamp duty, transfer fee, slippage) together with price-limit and suspension constraints (no fill when the limit is sealed; no fill without a quote). Paid tier (Standard and above). - New quant strategy list:
/v1/strategieslists every built-in strategy with its parameter ranges — 20 single-stock timing strategies (buy & hold / moving average / momentum / reversal / low volatility / volatility target / MACD / KDJ / RSI / Bollinger / WR / OBV / CCI / SAR / turtle / grid / DCA …) plus 9 portfolio selection strategies (momentum / reversal / low volatility / MA cross / volume breakout / small cap / large cap …). Paid tier (Standard and above); reading the list does not consume backtest quota. - New quant factor library:
/v1/factorslists all 88 stock-selection factors (code / Chinese name / unit / description / category) plus 22 presets and 8 usage rules. Factors feed directly into/v1/screenexpressions such asintersect([PE_TTM > 0, PE_TTM < 20, ROETTM > 15]). - Factors support historical cross-sections:
/v1/screen?date=YYYY-MM-DDreturns factor values for any past trading day (data exists from 2016-08-08), so you can chain factor → screen → backtest. - New quant playbook library:
/v1/playbooksgives a set of reusable "how to judge the market" workflows — each spells out its signals (formula + data + threshold), stages (e.g. sentiment: freeze / start / ferment / climax / ebb, with the action for each), what losses it avoids, and the measured evidence behind it (entries marked "pending" have not been run — we do not invent figures). 13 playbooks across 12 themes: risk control / trend / fundamentals / valuation percentile / rebalancing / money flow / market timing / sector rotation / sentiment cycle / volume structure / event-driven / limit-up chasing / ST restructuring. Paid tier (Standard and above). - "Avoid ST" and "trade ST" coexist: one playbook teaches how to screen ST names out, another how to hunt for turnarounds among them — the same stock looks like risk to a conservative investor and opportunity to an aggressive one, so we ship both. ⚠️ The measured result for screening ST out is counter-intuitive: momentum strategies do slightly worse without ST, small-cap strategies earn less but draw down less — it is a trade of return for drawdown, a risk-preference choice rather than a simple right or wrong.
- Two backtest modes:
mode=singlecomputes one stock live (milliseconds);mode=portfolioreads a precomputed multi-stock result (portfolio backtests touch the whole market and take much longer, so they are computed ahead of time). - Endpoint count 34 → 38.
- Also in Chinese: 中文更新日志.
2026-10-07EndpointsNew full-history K-line endpoint /v1/kline-full · 26 MCP tools- New full-history K-line endpoint:
/v1/kline-fullreturns 10 years of daily bars (from 2016-08-08, about 2,465 rows) with an optional adjustment modefq—qfqforward-adjusted /hfqback-adjusted /nofqraw. For backtests usehfqornofq: the forward-adjusted series is recalculated whenever a new ex-dividend occurs, so it is not reproducible. Paid tier (Standard and above). - How it differs from K-line:
/v1/klineis the lighter one (up to 1,212 rows, adjustment fixed to forward)./v1/kline-fullgives the full history plus a choice of adjustment, but A-share stocks only — for ETFs, convertible bonds, indices, sectors, Hong Kong and US use/v1/kline. - Endpoint total 33 → 34: still 7 free, now 27 paid.
- New MCP tool
ashare_kline_full: tool total 25 → 26. - Also in Chinese: 中文更新日志.
2026-10-07ReferenceFour new Wiki pages: response shapes, string numbers, batch fetching, cross-market differences- Response shapes:
datatakes four shapes (row arrays, table lists, pre-rendered text, multi-segment structures) — which of the 33 endpoints uses which, how to tell at a glance, and which key to read for programmatic access. - String numbers: most endpoints return prices and turnover as strings (
"1723.50") rather than numbers — which endpoints are affected, how to convert safely, and the traps when sorting or comparing. - Batch fetching: which endpoints accept several comma-separated codes at once, how to loop the ones that do not, and how to handle rate limits while looping.
- Cross-market differences: how the same endpoint differs in fields across A-shares / Hong Kong / US / Beijing, and which endpoints cover A-shares only.
- Available in both languages: 中文 Wiki.
2026-10-06EndpointsNew intraday endpoint /v1/minute · K-line gains minute periods- New intraday endpoint:
/v1/minutereturns the tick-by-tick trend (one price plus cumulative volume per minute) —days=1for today,days=5for the last 5 trading days. Paid tier. - K-line gains minute periods:
periodnow acceptsm1m5m15m30m60m120in addition to daily / weekly / monthly (window up to the last month), available on paid tiers. - Endpoint total 32 → 33: still 7 free, now 26 paid.
- New MCP tool
ashare_minute(for the intraday endpoint above): tool total 24 → 25; theashare_klinedescription now covers minute periods and date ranges too. - Anonymous rate change: after solving a PoW challenge the anonymous quota is now 15 requests/minute (was 60). Without a challenge it stays 5/min. Free data volume is unchanged.
- Also in Chinese: 中文更新日志.
2026-10-05ReferenceNew Wiki: field reference and convention notes- New Wiki reference section: field references, convention notes and common misuses — shorter than the tutorials, something to look up.
- K-line field reference: the close price field is called
lastnotclose, every numeric value is a string, andvolumeis in lots rather than shares. - Unit reference:
volumeis always lots;amountis CNY on K-line but CNY-10k on/v1/snapshot— mixing them up is off by 10,000x. - Adjustment convention: K-line prices are fixed to forward adjustment; there is no
adjustparameter, and no gap on ex-dividend dates. - Response envelope: what
ok,endpoint,tier,elapsed_ms,sourceanddataeach mean, and which responses do not use this structure. - Code format reference: how to write the
sh/sz/bj/hk/usprefixes (lowercase only), and what ETF, convertible bond and sector codes look like. - Data freshness: the data date lives in the
datefield, not your clock — markets and endpoints refresh at different cadences. - An empty result is not a failure: an empty array usually means a malformed code rather than a broken API, and empty results do not consume quota.
- Available in both languages: 中文 Wiki.
2026-10-04FixDSH plugin fix: tools no longer missing after install- Fixed the DSH plugin showing no tools after install: with 0.1.0 the plugin entry failed to activate, so none of the tools appeared. 0.1.1 is out — reinstall and it works, no config changes needed.
- How to tell if this affected you: if startup showed
1 entry did not activateorTypeError: Invalid effect, it was this. - How to reinstall: in Desktop / Web UI open the “Plugins” page and enter the package name
ashareapi-dsh-plugin(uninstall first if you already have it).
2026-10-03DataAnonymous quota now metered by rows: 250 per request / 100k per day- Anonymous (no-key) access is now metered by rows: up to 250 rows per request (list endpoints such as K-line) and up to 100,000 rows per day — plenty for day-to-day watching and small scripts.
- Keyed calls are unaffected: still up to 1212 rows per request with no daily cap. For bulk history pulls and backtests, just use a key (get an API key).
2026-10-02ReleaseDeepSeek Harness plugin released: install by package name- New DSH plugin page: a DeepSeek Harness plugin — once installed, DSH can query A-share data directly: quotes / K-lines / financials / money flow / top-trader boards / sectors / factor screening, 24 tools in total.
- How to install: in Desktop / Web UI open the “Plugins” page and enter the package name
ashareapi-dsh-plugin; or from the command line runnpx @deepseek-ai/dsh plugin --profile web add ashareapi-dsh-plugin. No config files to edit. - Free tools need no key: quote snapshot / K-line / hot list / market overview / breadth work right after install; the paid tools unlock with a key.
2026-10-02ReleasePi package released: one command gives Pi A-share data- Published on pi.dev (npm
ashareapi-pi): install withpi install npm:ashareapi-pi, then/reload— no JSON config, no scripts. - Covers 32 endpoints: quotes / K-lines / level-2 order book / financials / money flow / top-trader boards / sectors / convertible bonds / factor screening / macro — 5 of them free, no Key needed.
- Same source as the Agent Skill: one endpoint reference — Pi knows which endpoint to call, what parameters to pass, what units the fields use, and how to handle errors.
2026-10-01ReleaseSource opened: Python SDK / Node SDK / MCP server / Agent Skill on GitHub- All four repositories are open source (MIT): read the source, self-host, or file an issue — Python SDK · Node SDK · MCP server · Agent Skill.
- Official SDK updated to 0.1.9 (PyPI · npm): adds an English README.
2026-10-01GuidesNew guide: realtime quote and daily close with /v1/quote (no Key)- New Realtime quotes and daily close with /v1/quote: get the last price, OHLC, volume, turnover value and turnover rate in one GET — free, no Key, no signup. Covers unit conversion (volume in lots / amount in CNY / turnover as a percent), the data-date semantics behind "realtime vs close", batching and rate limits (anonymous 5/min, up to 60/min with one PoW challenge), and quote vs kline.
- Data from live responses (2026-10-01: 600667 returned 30 rows,
data[0]date=2026-09-30 last=17.29).
2026-09-30FixK-line count is capped at 1212; Retry-After added to rate-limit responses/v1/klinenow has a maximumcountof 1212: values above it return422. The default is still 30 — normal calls are unaffected.- Rate-limit (429) and data-source failure (502) responses now carry a
Retry-Afterheader, so clients can back off automatically instead of retrying blindly. No change needed for official SDK users — the SDK already backs off exponentially.
2026-09-30GuidesFour new practical guides: block trades / shareholder counts / K-line indicators / market breadth- New Reading A-share block trades: discount rate, institutional seats and both sides, with a real Hengrui 252m / -8% trade.
- New Reading shareholder counts: three tables, chip concentration, with the Tai Ji count-tripling case.
- New Technical indicators from K-line: MA / MACD / RSI / KDJ / BOLL in pure pandas (no TA-Lib).
- New Market temperature from breadth: advance ratio and limit-ups, tracking days for a turning point.
- All based on live data (2026-09-29), runnable as-is.
2026-09-29GuidesNew factor screening series: four guides (with a data health-check method)- New Factor screening basics: how to use
/v1/screen, how to read results, and the mandatory 3-step data health check (multi-metric consistency / ex-non-recurring / leverage) - with a real case showing how a screened PE 0.36 / ROE 883% was verified as an unusable factor value. - New Expression syntax & factor dictionary:
intersect([...])structure, 6 operators, unit and percentage rules, 85 verified factors, 14 copy-paste templates and 8 pitfalls. - New 22 ready-made screening strategies: grouped by valuation / profitability quality / technical / money flow, with live results from five strategies.
- New Value screening in practice: four-condition screen, a four-step investigation, then turning it into a reusable function that caught the outlier without false positives.
- All examples are live responses (reproduced 2026-09-29) and can be run as-is.
2026-09-28FixData source count basis corrected: 67 → 70- We aggregate 70 data sources (some pages previously showed 67) — the 67 came from an earlier per-category estimate, while 70 is counted from the sources we actually integrate.
- What changed is the displayed number, not the data: responses, fields and the endpoint count (32) are unchanged; affected pages — the homepage status card, the status page, pricing, comparison pages and the footer — now read 70.
- The status page lists the size of each category: quotes / financials / money flow / top-trader boards / sectors / attention / filings & news / other — the groups add up to 70.
2026-09-28FixK-line basis: fixed to forward-adjusted (no need to adjust yourself)/v1/klinereturns forward-adjusted prices — no gap on ex-dividend days; there is noadjustparameter (the basis is fixed, not configurable). ⚠️ Please do not adjust again (double adjustment leaves prices that no longer match the real market); unadjusted / back-adjusted series are not provided.- When our prices do not match an unadjusted source, the difference lands exactly on windows that cross an ex-dividend date (more visible on long windows such as MA120), while short windows (MA5 / MA10) usually agree — that is a basis difference, not a data error.
- This basis is now stated everywhere you read us: guides and FAQ · endpoint descriptions (
/openapi.jsonand the endpoint pages) · the MCP toolashare_kline· the Agent Skill · both official SDKs (thekline()docs and READMEs for Python and Node.js). - The MCP server address is
https://api.ashareapi.com/mcp(use this one in your client);ashareapi.com/mcpis the configuration guide page, not the endpoint. - New comparison page: ashareapi vs BaoStock — a free historical database vs a hosted service: BaoStock switches between three price-adjustment modes via
adjustflagand exposes adjustment factors, plus 5/15/30/60-minute bars; we provide real-time snapshots, structured money-flow surfaces (money flow / top-trader boards / sectors / convertible bonds) and native AI Agent access. Includes "who each fits" and honest boundaries.
2026-09-26DataField rename: turnover rate on /v1/quote and /v1/kline is now turnover (was exchange)- Unified field name: the turnover-rate field on
/v1/quoteand/v1/klineused to beexchange— the raw column name of the data source, which reads like "exchange" the venue (our own endpoint description was misled by it and labelled the field as the exchange name while claiming "no turnover rate" — see the fix in the entry above). It is nowturnover, the same name and value as/v1/snapshot(sz000001=0.54in both) - ⚠️ Breaking change: code reading
row["exchange"]must switch torow["turnover"]. This is a direct replacement — noexchangealias is kept (rationale: no external customer depends on this field yet) - Propagated to: HTTP endpoint description · OpenAPI (
/openapi.json) · MCP tool descriptions (ashare_quote/ashare_kline) · endpoint pages · guide code samples (Python / Node.js) · Agent Skill - Also fixed a wrong statement in the guides: it said the turnover rate lives on
quote/snapshotand not on kline —/v1/klinedoes return it - Unchanged:
/v1/snapshotalready usedturnover. The/v1/klineendpoint description previously did not mention the field at all — now it does
2026-09-26FixFixed: /v1/quote turnover field was mislabelled as "exchange"- The
exchangefield IS the turnover rate (%) — not the exchange name: an earlier description labelled the 8th field as "exchange" and claimed "no turnover rate" — a conclusion drawn from the field NAME alone. Measured values are numeric (sz000001=0.54,sh600036=0.24) and identical to/v1/snapshot'sturnover(0.54 for sz000001 in both) → it is the turnover rate - Cross-check: 1,043,819 lots ÷ 19.406e9 shares = 0.538% ≈ the measured 0.54; the same-day change (−0.44%) and amplitude (1.59%) are both ≠ 0.54 (elimination)
- The description now reads "lightweight, 8 fields: last price / OHLC / volume / amount / turnover rate" and notes the field name is
exchange(counter-intuitive — it is NOT the exchange name) - Also unified
/v1/snapshot's field count to 35 fields (the HTTP side said "24+", the MCP side said "35")
2026-09-26FixAgent Skill doc fix: wrong response-shape description (copying it caused KeyError)- Corrected the "unified envelope" claim: the skill said every response carries
structured/tables. Measured across all 32 endpoints, they are added only whendatais Markdown text (9 endpoints:market-overview/changedist/lhb/sector/sector-valuation/bond/etf/screen/macro). Other endpoints have no such keys, so AI code copyingbody["structured"]hits a KeyError (usebody.get("structured")) - Added a "four shapes of
data" quick-reference: object array (most endpoints) / table list (finance) / sectioned structure (shareholder,calendar) / Markdown text — plus a shape-agnostic row extractor - Documented response shapes:
finance= table list ·shareholder/calendar= sectioned (with section keys) ·fund= flat dict; fixedprofilebeing labelled "returns a string" (it is an object array) /v1/snapshotfield count: 24+ → 35 fields (now consistent with the endpoint page and the MCP side)- Skill version 0.2.0 → 0.2.1; re-packed both the
/skilldownload and the Agent Skills distribution zip (/.well-known/agent-skills/)
2026-09-25FixFixed: /v1/calendar returned stale data + used the wrong data channel- It returned months-old stale data: /v1/calendar used a crippled channel whose output had only two columns,
date+event.zy(the latter always "1", with no business fields at all), and it ignored thedate/limitparameters (passingdate=2026-09-25still returned the full ascending list from 06-01) — so clients received data from months ago, the opposite of a "calendar" - Switched to the correct channel: it now returns sections by event type (earnings disclosure / dividend / IPO / meeting / lockup release / rights issue) with business fields such as
symbol/stockName/registrationDate/exDividendDate/paymentDate, anddatenow takes effect (defaults to from today, returning upcoming events over roughly the following month) - Response structure change:
datais now a sectioned structure —data.tablesis an ordered list of sections, each withtitle(Chinese label) /slug(English key) /rows; plusdata.data{slug: rows}as a convenience index - Slugs:
financial_report/dividend/ipo/meeting/lockup_release/rights_issue(plustrading_halt) - ⚠️ Breaking change: code that iterated over
datato read rows must now go throughdata.tablessection by section (see the endpoint docs example) - Also corrected the parameter docs:
dateis the window start date (blank = from today);limitis a data-source scale parameter, and per-section counts are not exactly this value
2026-09-25FixFixed: /v1/shareholder dropped holder sections + response structure change- Only the first section was returned: the data source returns three sections (A-shares: top-10 holders / top-10 float holders / holder counts; HK: shareholder info / holder distribution / institutional holdings), but the parser kept only the first table — the other two were silently dropped
- Response structure change:
datais now a sectioned structure —data.tablesis an ordered list of sections, each withtitle(Chinese label) /slug(English key) /rows; plusdata.data{slug: rows}as a convenience index - A-share slugs:
top10_holders/top10_float_holders/holder_count(includestotalSHNum/avgHoldSharesfor chip concentration) - HK slugs:
shareholders/holder_distribution/institutional_holdings(semantically different from A-shares — do not index by position) - ⚠️ Breaking change: code that iterated over
datato read rows must now go throughdata.tablessection by section (see the endpoint docs example) - Also fixed the
okflag: when the data source has no data (e.g. an index code) it used to returnok:truewith an empty list — it now correctly reportsok:false - Also documented the response shape of
/v1/quote:datais an array (30 daily bars by default,data[0]is the latest). This was previously unstated, so parsing it as an object would fail — read the current price fromdata[0].last - Also documented the response shape of
/v1/quote:datais an array (30 daily bars by default,data[0]is the latest). This was previously unstated, so parsing it as an object would fail — read the current price fromdata[0].last
2026-09-25FixFixed: /v1/finance period direction + response structure change- Period direction fixed:
num=Nused to return the earliest N periods (possibly excluding the newest disclosure) — it now returns the latest N periods, matching the parameter description - Response structure change:
datawent from a list of periods to a list of tables —data[0]=income ·data[1]=balance sheet ·data[2]=cash flow; each table is[{EndDate: …, field: …}]in descending order (latest period first) - Far more fields: the three tables now carry 85 / 70 / 31 columns (full Tushare-style line items); the previous single table had only 36
- ⚠️ Breaking change: code that iterated over
datamust index by table first (see the docs example). Note thatstructuredwas always empty for this endpoint — it had been falling back todata - Also fixed /v1/block-trade: it used to return only meta info and dropped the trade details — it now returns the full detail rows (price / premium-discount / volume / both-side brokerages)
2026-09-23EndpointsNew endpoint: full quote profile (/v1/snapshot)- One request returns the full picture: price + order book + valuation (PE TTM/dynamic/static + PB) + market cap + shares + limit up/down + volume ratio + amplitude + speed
- Split from /v1/quote: quote is light (8 fields, free); snapshot is full (24+ fields, pro tier) — use quote when you only need the price
- A-shares only (HK/US field layouts differ; parsing them would return wrong data, so they are not supported — use quote)
- Endpoint count 31 -> 32; MCP tools 23 -> 24 (ashare_snapshot)
2026-09-23DataOrder-book fields flattened (b1_p / a1_v …)- Field names follow the common data-interface convention: b1_p/b1_v ~ b5_p/b5_v (bids), a1_p/a1_v ~ a5_p/a5_v (asks)
- Flat layout (one row, 30 columns) instead of nested arrays → read df["b1_p"] directly, no df["buy"][0][0]
- Response data is now always a row array → both SDKs (Python / Node) return DataFrames / arrays of objects automatically
- SDKs bumped to 0.1.2; limit-down zeros the bids and limit-up zeros the asks (normal, documented)
2026-09-23EndpointsNew endpoint: five-level order book (/v1/orderbook)- Bid 1–5 / Ask 1–5 prices and resting size (lots), plus last / change / data time
- Second-level snapshot (10s cache) · meaningful intraday · after close it is the final snapshot of the day
- Endpoint count 30 → 31; MCP tools 22 → 23 (ashare_orderbook)
- Both SDKs (Python / Node) gain an orderbook() method
2026-09-23ReleaseNode.js / TypeScript SDK published on npm- npm install ashareapi — 30 endpoints as 30 methods (npmjs.com/package/ashareapi)
- Zero runtime dependencies (native fetch, Node >= 18); ESM + CommonJS builds with TypeScript types
- Free endpoints need no key; returns arrays of objects; both camelCase and snake_case method names work
- Tutorial: /docs/guides/nodejs-ashare-quotes
2026-09-23ReleasePython SDK published on PyPI- pip install ashareapi — official Python SDK, 30 endpoints as 30 methods (pypi.org/project/ashareapi)
- Free endpoints need no key; returns pandas DataFrames (falls back to list[dict] without pandas)
- Symbol formats: sh600667 / 600667.SH / bare 6 digits all work
- 5 exception types separate "no data right now" from "fetch failed" (EmptyResultError / UpstreamError)
- Tutorial: /docs/guides/python-sdk
2026-09-23ReleaseAgent Skill launched (works with 45+ clients)- Drop one folder into Claude Code / Cursor / Codex / OpenCode and 45+ other Agent Skills clients — the AI then knows how to use the API
- Open standard (agentskills.io) · only the 6 standard frontmatter fields, so it stays portable across clients
- Progressive disclosure: an always-on SKILL.md plus 4 references (endpoints / fields / errors / SDK) read on demand
- Page and download: /skill
2026-09-21EndpointsAnonymous quota boost + 29 endpoints- New endpoint /v1/challenge: solve one challenge to raise the anonymous quota from 5/min to 60/min
- Endpoint count 28 → 29 (26 data endpoints + 3 utility endpoints: health / challenge / usage)
- The hot list limit now applies (/v1/hot, default 30, max 50)
- CORS enabled — the status page and browsers can call the API directly
2026-09-16ReleaseService launch- All 28 endpoints open (6 free + 22 paid)
- Seoul node · multi-source fetch (100% success / 862ms)
- HTTPS site-wide · certificates auto-renew
- MCP access (23 tools)
- IP-level anti-resale protection
2026-09-15DataData source expansion- 67 data sources · picks the fastest, skips the ones that keep failing
- Three money-flow sources back each other up
- Split of top-trader boards (institutions / hot money / seats)