Best Market Data APIs for AI Agents, LLMs and Trading Bots in 2026
Market data APIs compared for AI agents, LLMs and trading bots: which vendors run a hosted MCP server and how it signs in, who publishes llms.txt, a skill file or an OpenAPI spec, rate limits an agent loop can live with, coverage beyond prices, and pricing read on each vendor's page on October 13, 2026.
For an AI agent, the best market data API is the one the agent can discover and call without custom glue: a hosted MCP server it can connect to, a published llms.txt or skill file it can read, JSON it can branch on, and a rate limit sized for an agent's bursts. On those axes, as of October 13, 2026, ten of the fourteen providers below run a vendor MCP server, eleven publish a real llms.txt, and the free tiers range from 25 requests a day to 200 calls a minute. The data behind the server then decides the pick: prices and history (Massive, Twelve Data, Alpaca), statements and filings (Financial Datasets), options flow (Unusual Whales), congressional and alternative data (Quiver), and the market's read on a stock with the smart money behind it (SentiSense).
Every price and limit below was read on the vendor's own page on October 13, 2026. Items we could not confirm that day are marked unverified. SentiSense is our product; its row is graded as hard as the others and its shortfalls are listed.
This post compares the data providers. If you want to connect a chat app to one of their servers today, the best MCP servers for stock market data walks through that; if you are choosing a general stock data API for a script, the best stock data APIs is the wider comparison.
What makes a market data API good for an AI agent?
Five things, none of which appear on a feature grid built for humans.
- A hosted MCP server, and how it signs in. Hosted means the vendor runs it and the client connects by URL; local means you install a package and run it with your key. OAuth sign-in lets a non-developer connect from Claude, ChatGPT or Grok without pasting a secret; an API key in the URL works but ends up in logs and chat histories.
- Agent-readable docs. An
llms.txtat the domain root, a skill file the agent can install, or an OpenAPI spec. Without one of these the model writes calls from memory, and memory is where endpoint hallucinations come from. - JSON the model can branch on. Stable field names, explicit nulls for "no data", a distinct error code for rate limits versus quota, and a
Retry-Afterheader. Preview or truncation markers matter too: an agent must know when it received a free slice rather than the full record. - Rate limits that fit a loop. A human makes a few calls a minute; an agent fans out. A daily cap of 25 is one research session. Per-minute limits with no monthly cap suit agents best.
- Coverage beyond prices. Prices are commodity and every provider has them. Sentiment, filings, insider and congressional trades, 13F holdings, options positioning and earnings are what let an agent answer "why" rather than "how much".
How do the providers compare on agent access?
| Provider | MCP server | Sign-in | Agent-readable docs | Free limit |
|---|---|---|---|---|
| SentiSense | Hosted, vendor-run, 11 read-only tools; in Claude's connector directory | OAuth | llms.txt, skill.md, .well-known/agent-skills, 20+ installable skills (npx skills add SentiSenseApp/skills); no OpenAPI file |
30/min, 1,000/month |
| Massive (formerly Polygon.io) | Hosted, vendor-run (mcp.massive.com); local via pip |
OAuth | llms.txt; skill files mentioned without a URL; no OpenAPI found |
5 calls/min, end of day |
| Alpha Vantage | Hosted, vendor-run (about 120 tools); local via uvx; README cites Claude and ChatGPT directory listings | OAuth recommended; key in URL deprecated | No llms.txt (404); no OpenAPI or skill found |
25 requests/day |
| Finnhub | Hosted, vendor-run, 70 tools | OAuth | Swagger spec; llms.txt URL serves an HTML page |
60 calls/min |
| Twelve Data | Hosted (mcp.twelvedata.com); local via Python or Docker |
OAuth (hosted); key (local) | llms.txt plus per-endpoint Markdown, OpenAPI, a ClawHub skill |
8 credits/min, 800/day |
| EODHD | Hosted, open source; v1 and v2 | v1 API key in URL; v2 OAuth 2.1 | llms.txt; a Claude plugin covering 72 endpoints; no OpenAPI found |
20 calls/day, 20/min |
| Financial Datasets | Hosted (mcp.financialdatasets.ai); local repo |
OAuth (hosted); key (local) | llms.txt, openapi.json, an agents.md guide with agent self-signup |
No free tier; limits not stated |
| Financial Modeling Prep | Hosted, vendor-run; calls count against the API limits | API key in URL | A 567-byte llms.txt; OpenAPI unverified (docs pages return 403 to fetchers) |
250 calls/day (unverified) |
| Unusual Whales | Hosted, vendor-run, 74 tools; local via npm | API key, or account sign-in from ChatGPT, Claude and Grok | llms.txt, skill.md, OpenAPI |
No free tier; MCP 120/min, unlimited per day on paid plans |
| Quiver Quantitative | Hosted (mcp.quiverquant.com), 18 tools, via the mcp-remote proxy |
API key | llms.txt; no OpenAPI found |
No free tier; limits not stated |
| AxionQuant | Local only (npm i @axionquant/mcp), 100+ tools per the vendor |
API key | llms.txt, OpenAPI listed, SDKs |
1,000 calls/month; no per-minute limit stated |
| Alpaca | Local only (uvx or Docker); "does not configure MCP OAuth" | Key and secret | llms.txt; no OpenAPI or skill found |
200 calls/min |
| Databento | None found | n/a | llms.txt URL serves an HTML page; no spec found |
Not stated; $125 of free credits |
| Tiingo | None found | n/a | llms.txt |
50 requests/hour, 1,000/day |
Two notes. Unusual Whales' own guide to market data APIs for AI agents, which some answer engines cite, names three providers and says Polygon.io and Twelve Data have no MCP server; both now publish hosted ones. And a vendor's llms.txt is only useful if it is real: Finnhub's and Databento's URLs return an HTML page, and FMP's is 567 bytes.
What does each cost, and what is behind the server?
Prices checked October 13, 2026, on each vendor's pricing page. Entry plan is the cheapest paid tier that includes API access.
| Provider | Free tier | Entry paid plan | Coverage beyond prices |
|---|---|---|---|
| SentiSense | 1,000 requests/month | PRO $15/mo or $144/yr, no monthly cap, 300/min (Pro Max 1,500/min) | Model-scored sentiment and the SentiSense Score, the A to F Rating, news, insider, congressional, 13F, end-of-day options analytics, earnings, financials and KPIs, a screener, a company knowledge graph; US stocks and ETFs only |
| Massive | End of day, 2 years | Starter $29/mo (15-minute delayed) | Trades and quotes, options, forex, crypto, futures; fundamentals on Advanced ($199) or a $29 add-on. Sentiment, insider and congressional data: not seen on the pages checked (unverified) |
| Alpha Vantage | 25 requests/day | $49.99/mo, 75 requests/min | News sentiment, insider transactions, congressional trades, institutional holdings, options, earnings calendar, fundamentals |
| Finnhub | 60 calls/min, personal use | All-In-One $3,500/mo billed annually, personal use; the pricing page lists no tier between | Fundamentals, estimates, filings, insider transactions, social sentiment (premium), news |
| Twelve Data | 8 credits/min | Grow $79/mo or $790/yr by the plan card; a second table on the same page says "from $29/mo" (which applies is unverified) | Fundamentals, earnings, insider transactions, institutional and fund holders, EDGAR archive; no sentiment |
| EODHD | 20 calls/day | EOD Historical $19.99/mo; news sentiment on every plan | Fundamentals, insider filings and earnings from $59.99; congressional trades on the $99.99 plan (beta); US options a $29.99 add-on; SEC filings in beta; no 13F |
| Financial Datasets | None | $20 of credits (1,000 requests); Personal $200/mo | Statements, segments, KPIs, SEC filings with extracted sections, insider trades, 13F and activist holdings, macro; no options, sentiment or congressional data |
| Financial Modeling Prep | 250 calls/day (unverified) | Starter $29/mo; Premium $69; Ultimate $139 (3,000 calls/min) | Statements, filings, insider and congressional trades, earnings, news; 13F on Ultimate; social sentiment endpoints are legacy |
| Unusual Whales | None (7-day trial) | API Basic $150/mo; Advanced $375; personal use, business plans from $625/mo billed annually | Real-time options flow, dark pool, congressional and insider trades, institutional data, earnings, fundamentals, news |
| Quiver Quantitative | None | Hobbyist $30/mo ($25 billed annually), non-commercial | Congressional and insider trades, institutional holdings, lobbying, government contracts, patents, off-exchange prints; no options or sentiment |
| AxionQuant | 1,000 calls/month, US stocks and profiles | Starter $49.99/mo ($39.99 billed annually) | Filings, earnings, news on Starter; sentiment, ESG and credit ratings on Pro at $199.99/mo |
| Alpaca | IEX data, 15-minute delayed through the API | Algo Trader Plus $99/mo (all US exchanges, OPRA options) | Stocks, crypto, options, news; no sentiment, filings, insider, congressional or 13F data |
| Databento | $125 of credits | Standard $199/mo live | Raw exchange trades and quotes, every asset class; nothing derived |
| Tiingo | 1,000 requests/day | Power $30/mo ($300/yr) | End of day, IEX, crypto, forex; news and fundamentals as add-ons; no sentiment, filings, insider, congressional, 13F or options |
What do the rate limits mean for an agent loop?
Count calls per task, not per user. A watchlist of twenty names with four lookups each is eighty calls before the agent writes a sentence. On a 25-a-day plan that is one task; on 5 a minute it is sixteen minutes of waiting; on 60 a minute it is under two. Daily caps punish agents hardest because an agent does not spread its work across the day. Prefer per-minute limits with no monthly ceiling, and check how the vendor signals a limit: a 429 with Retry-After lets the agent wait and retry, while a quota exhaustion it cannot distinguish from a rate limit sends it into a retry loop that burns the rest of the month.
Credit systems need one more check. Twelve Data and EODHD charge different credits per endpoint (EODHD counts 1 for end-of-day, 5 for intraday or news, 10 for fundamentals), so the advertised limit is an upper bound on cheap calls, not a budget of lookups.
Where does SentiSense honestly sit?
Where it loses: no global equities, no tick data or streaming, prices delayed 15 minutes, no OpenAPI file (the reference is skill.md and llms.txt), and a free tier with a monthly cap of 1,000 rather than Finnhub's uncapped 60 a minute.
Where it wins for an agent: a hosted MCP server with OAuth, listed in Claude's connector directory and documented for Claude Code, Codex, ChatGPT, Grok and Cursor; eleven read-only tools whose free-tier responses carry an explicit preview marker so the agent knows what it did not get; agent-readable docs in three forms; 20+ installable skills with fixed output templates for research workflows; and the coverage that lets an agent answer "why" on a US stock: model-scored sentiment across news, Reddit, X, YouTube and Substack, the SentiSense Score and Rating, insider, congressional and 13F data, end-of-day options analytics ranked against each stock's own history, earnings and KPIs, and a screener. PRO is $15 a month list with no monthly cap at 300 calls a minute.
The honest pairing: a price feed from Massive, Twelve Data or Alpaca for the numbers, and SentiSense for the read on them. How to connect market data to your AI agent shows the four routes with a runnable loop.
Which one should an agent builder pick?
- A hosted MCP a non-developer can connect, with OAuth and no key to paste: SentiSense, Massive, Alpha Vantage, Finnhub, Twelve Data, Financial Datasets, or EODHD v2.
- Prices and history for a bot you are writing: Massive or Twelve Data; Alpaca if the bot also trades through Alpaca.
- Statements and filings the model can read section by section: Financial Datasets.
- Options flow in real time: Unusual Whales, on its personal-use terms; for end-of-day options context ranked against history, SentiSense on a free key.
- Congressional, lobbying and insider data: Quiver, or SentiSense if you also want sentiment and 13F flows under one key.
- The widest catalog on a budget: EODHD or Financial Modeling Prep.
- Raw exchange tape: Databento, with no MCP and the OPRA licensing that comes with redistribution.
Data is provided for research and education, not investment advice.
FAQ
Which market data APIs have an MCP server? As of October 13, 2026, vendor-run hosted MCP servers are published by SentiSense, Massive (formerly Polygon.io), Alpha Vantage, Finnhub, Twelve Data, EODHD, Financial Datasets, Financial Modeling Prep, Unusual Whales and Quiver Quantitative. Alpaca and AxionQuant ship local MCP servers you install and run with your own key. Databento and Tiingo publish none.
What is the best free market data API for an AI agent? It depends on the call pattern. For a hosted MCP connection with no key to paste, SentiSense's free account allows 1,000 calls a month at 30 a minute, and Finnhub's free key allows 60 calls a minute. Twelve Data's free plan is 8 credits a minute and 800 a day, Massive's is 5 calls a minute of end-of-day data, EODHD's is 20 calls a day, and Alpha Vantage's is 25 requests a day. Alpaca's free market data plan allows 200 calls a minute of IEX data. Financial Datasets, Unusual Whales and Quiver have no free tier.
Do I need MCP, or is a REST API enough for an LLM? A REST API is enough when you write the tool definitions yourself and run the loop in your own code. MCP removes that work: the vendor publishes the tools and their schemas, and any MCP client (Claude, ChatGPT, Cursor, Codex, Grok and others) discovers them. For an app you ship, use the REST API or SDK directly and read the vendor's llms.txt or skill file so the model writes correct calls; for an assistant a person talks to, a hosted MCP server with OAuth is the shortest path.