Prediction Market MCP Servers and AI Agent Tools (2026)

Search for prediction market AI agents and the top result by star count is Polymarket/agents: 3,771 stars, 823 forks, MIT licensed, and archived since its last push on 5 November 2024. It is the most starred project in the category and nobody has touched it in over a year. Polymarket's live answer is an Agent Skills repository rather than an MCP server, and that repository ships with no license file. Neither Polymarket nor Kalshi publishes an official MCP server, so nearly everything people connect to Claude or Cursor here is third-party code of uneven quality. This post covers what is maintained, what the free tiers really include, and which projects default to not trading.

Repository facts here come from the GitHub REST API: license field, stars, forks, the archived flag and last push dates. Pricing and tool lists come from each vendor's own pricing, docs and README pages. Tool counts are reported as claims, because READMEs were read but no source code was audited. One gap, stated openly: Polymarket's documentation hosts did not resolve during checking, so no Polymarket rate limit or endpoint figure appears here.

The flagship agent repo is archived

Polymarket/agents is described as a way to trade autonomously on Polymarket using AI agents. It is MIT licensed and archived, with a last push on 5 November 2024. Archived means read only: no merges, no fixes, no maintenance. Any tutorial built on it is building on a frozen base.

That matters more here than in most categories, because the rest of the Polymarket organization is busy. ts-sdk was pushed on 17 August 2026, py-sdk on 14 August 2026 and the v2 CLOB clients on 5 August 2026. The client tooling moved on and the agent framework did not.

MCP servers are only half the picture

Treating "prediction market AI agent" as a synonym for "MCP server" is already out of date. Polymarket's current answer is Polymarket/agent-skills: 181 stars, created and last pushed on 19 February 2026, license field null. It ships a web3-polymarket directory with SKILL.md plus reference files for authentication, order patterns, market data, WebSocket, CTF operations and gasless transactions.

The design reason is stated in the repo: progressive disclosure, with roughly 200 lines of context loaded up front and roughly 1,700 lines of reference pulled in on demand. That keeps the agent's context small until it needs the detail, a different trade to an MCP server exposing a fixed tool surface at connect time. The caveat runs through this category: no license file means no grant.

Skills are not a Polymarket quirk. Myriad's CLI ships three, for market discovery, trade execution and wallet operations, Simmer ships what it describes as 28 skills, and newyorkcompute/kalshi offers Agent Skills alongside its MCP server.

Which venues ship agent tooling themselves

Only three of the venues checked ship anything agent-shaped from the exchange side.

Myriad goes furthest. Its builders page lists an HTTP API, a JavaScript SDK, an MCP server for LLMs and skills for agent frameworks, and says the Myriad CLI is the fastest way to get an agent trading there. That CLI is MyriadProtocol/myriad-cli, MIT licensed, published as npm install -g @myriadmarkets/cli, and myriad mcp starts it as an MCP server. It covers discover, decide, execute and claim, and has a dry run flag.

Limitless publishes a dedicated page on building a trading agent, pointing at the TypeScript template limitless-labs-group/agents-starter with three built-in strategies: cross-market market making, oracle arbitrage and a certainty closer. The docs say you can fork it, set two environment variables and run a strategy in dry run within five minutes. The template has 7 stars and no license file, so the open source label is not backed by a grant. Limitless also has the deepest official SDK coverage of any venue checked, across four languages, but programmatic access is gated behind an application form.

Kalshi, SX Bet, Opinion and Predict.fun ship none. Kalshi's docs index covers getting started, API reference, WebSockets, FIX and SDKs with no MCP entry, and its official SDKs are kalshi_python_sync, kalshi_python_async and kalshi-typescript. Opinion has a full REST API, WebSocket channels and two SDKs, but no agent page, and Predict.fun publishes no API reference or SDK at all.

Hosted MCP services, and what the free tiers leave out

Four hosted services compete here, and two have free tiers that exclude the exact capability the product is sold on.

Parlay is the sharper trap. Free is $0 a month for 15 calls with only search_markets and market_brief, roughly one call every two days. Every cross-venue tool sits on Pro at $29, which raises the allowance to 150 calls and opens all six tools. Since Parlay's stated purpose is cross-venue comparison, the free tier cannot do the thing the product is for. Note also that parlay-run/parlay-mcp is MIT but is only the connection reference for the hosted service, which is itself PMXT-backed.

DepthFeed is blunter. Its docs state the free Explorer tier streams Polymarket and Binance BTC, and that Kalshi and Limitless channels require a paid plan. Ask for them on Free and you get a 402 with the code ENDPOINT_NOT_IN_PLAN, and a non-BTC coin returns 402 COIN_NOT_IN_PLAN. The product is 11 read-only tools across three venues and 7 assets, so free is one venue and one asset. Separately, kalshi-mcp.com is a paid Kalshi service at $5, $19 and $59 a month with only a 7 day trial, and states on its own page that it is unofficial.

The free path that actually works

You can assemble a capable read-only setup without paying anyone. JamesANZ/prediction-market-mcp is MIT, 36 stars, last pushed 16 January 2026, covering Polymarket, PredictIt and Kalshi with market data and calculated odds, no keys required.

For Kalshi, 9crusher/mcp-server-kalshi is the best designed free option: MIT, 25 stars, last pushed 28 July 2026. Credentials are optional, because all market and rules tools work unauthenticated and only portfolio and order tools need a key. Its standout tool extracts the text of the actual legal contract PDF so the agent can read the resolution terms. Its README also flags a real constraint: Kalshi has no free-text search, so the list_markets filters are the search.

For multi-venue work without a hosted bill, pmxt-dev/pmxt-mcp is MIT, published as @pmxt/mcp on npm, and runs in two modes. Set an API key and it calls the hosted API; set none and it assumes you are running the PMXT core server locally, which needs no key at all. To practise safely, agent-next/polymarket-paper-trader is MIT with 372 stars, starting an agent with $10,000 of paper money against live Polymarket order books, with level-by-level execution and an explicit fee model.

Third-party servers, ranked by activity rather than stars

The most starred active third-party Polymarket MCP server is caiovicentino/polymarket-mcp-server: 643 stars, MIT, last pushed 30 July 2026. Its README claims 45 tools across discovery, analysis, trading, portfolio and real-time categories, with configurable safety rails for order size and exposure. Those are README claims, not audited counts.

Two smaller active MIT alternatives: BrainDAO/mcp-polymarket, 19 stars, pushed 10 August 2026, and demwick/polymarket-agent-mcp, 15 stars, claiming 48 tools spanning trading, smart money tracking, copy trading and risk management. For a multi-venue harness, SpartanLabsXyz/simmer-sdk is MIT with 48 stars and a push on 17 August 2026, bundling skills, an MCP server published as simmer-mcp on npm, and a Python SDK, with the venue set by an environment variable across a simulator, Polymarket and Kalshi. It also documents a real limit: unfiltered browsing is windowed to the newest 1,000 or so markets.

Then the ones directories still list that you should skip: berlinbra/polymarket-mcp has 142 stars but was last pushed in June 2025, and shaanmajid/prediction-mcp is archived. newyorkcompute/kalshi carries a different trap. The repo is MIT and was pushed in August 2026, claiming 14 MCP tools plus Agent Skills, while its npm package has not been updated since December 2025.

Safety rails to check before you connect a wallet

An agent with order tools and a private key can lose money faster than you can read the transcript. The projects worth trusting all default to not trading. 9crusher's Kalshi server targets the demo sandbox unless you set the environment to production, and its order tools return a preview unless you pass a confirmation flag. Simmer's README instructs that skills default to dry run and require an explicit live flag. PMXT's order tools require explicit confirmation, and Myriad's CLI has a dry run flag.

Two structural risks are easy to miss: self-custody venues need a wallet private key in your environment, a very different posture to an API key you can scope and revoke, and Limitless has no sandbox at all.

Rate limits and data windows your agent will hit

Agents burn request budgets far faster than humans, so plan against the published numbers. Kalshi uses a token bucket with separate read and write buckets, where the default cost is 10 tokens per request. Tokens per second run from Basic at 200 read and 100 write, through Premier at 1,000 and 1,000, up to Prestige at 10,000 read and 8,000 write. Advanced and above can burst to twice the per-second budget, and exceeding a bucket returns a 429 with no cooldown penalty.

Myriad publishes 200 requests per second with a key, 30 per 10 seconds without one, and 200 orders per 10 second window per trader address, with legacy bare API keys sunsetting on 15 August 2026. Opinion is a flat 15 requests per second with no tiers. SX Bet publishes limits per endpoint, including 5,500 order posts and just 60 cancel-all calls per 60 seconds. History is the other constraint: Kalshi targets a three month window for live data, after which you move to separate historical endpoints. It serves real-time market data without authentication, while fills, orders and positions need auth.

Where a cross-venue aggregator fits

Predictefy ships an MCP server of its own, @predictefy/mcp, run with npx -y @predictefy/mcp and MIT licensed. It exposes read and intelligence tools across 15+ venues, with trade tools kept behind a separate guardrail and a hard rule that no single tool both builds and submits an order. TypeScript and Python SDKs cover the same surface. All three are on beta versions, so pin an exact release rather than tracking latest.

What it does solve is the coverage problem on the human side. DepthFeed covers three venues, Parlay four, and most third-party MCP servers cover one or two. Predictefy aggregates Polymarket, Kalshi, Limitless, Opinion, Myriad and the other venues it tracks into a single browser view with an arbitrage terminal, which is a different job to giving a model tool access. The sensible split is an MCP server for the venue your agent trades on, and an aggregator for the cross-venue picture. One disclosure in keeping with this post: the arbitrage API is free to use.

Conclusion

The category is younger than its star counts suggest. The most visible project is archived, the official skills repo has no license, and two of the four hosted services put their headline capability behind a paywall while still calling the entry tier free. The practical path is unglamorous: start with a keyless MIT server, practise in the Polymarket paper trader, run PMXT's server locally if you need several venues, and pay only once you know which tools your agent calls. Check the license, the archive flag and the last push date on anything a directory recommends.