metaculus-mcp

metaculus-mcp

A remote MCP server wrapping the Metaculus API that enables agents to browse forecasting questions, submit and withdraw forecasts, and post rationale comments using user-provided API tokens.

Category
访问服务器

README

Metaculus MCP Server

A remote MCP server for the Metaculus forecasting platform, running on Cloudflare Workers (free tier, stateless, zero server-side secrets). It wraps the official Metaculus API v2 so any MCP-capable agent (Claude, etc.) can browse forecasting questions, submit/withdraw forecasts — including continuous questions via an automatic percentile-sketch → 201-point-CDF builder — post rationale comments, and pull its own score/forecast CSV exports.

Every user supplies their own free Metaculus API token via a request header. The Worker holds no credentials and forwards the token only to https://www.metaculus.com.

Why this exists: Metaculus runs bot-only AI Forecasting Benchmark tournaments (metaculus.com/aib/) and publishes its own bot-template repos, yet as of 2026-07-12 there was no Metaculus MCP server anywhere — 0 hits in registry.modelcontextprotocol.io, 0 among 127 Metaculus-related GitHub repos — while adjacent platforms (Polymarket, Kalshi) have 10+/6+ servers each. This fills that hole.


Tools

Tool What it does API endpoint
list_questions Browse/filter the question feed: status, type (binary/numeric/date/discrete/multiple-choice/conditional/group), tournament and category slugs, time windows, ordering (e.g. -hotness). Returns compact summaries. GET /api/posts/
get_question Full detail for one post: description, resolution criteria, fine print, subquestions, and the question.scaling data needed to build continuous forecasts. GET /api/posts/{postId}/
get_my_predictions Posts a given user (you) has forecast on — your active book / track record surface. GET /api/posts/?forecaster_id=…
submit_forecast Submit one or many forecasts: probability_yes (binary), probability_yes_per_category (multiple-choice), a full continuous_cdf, or a percentiles sketch that the server converts into a spec-valid CDF (linear & log scaling, open/closed bounds, minimum-slope rules handled per the spec's documented algorithm). POST /api/questions/forecast/
withdraw_forecast Withdraw current forecasts on one or many questions. POST /api/questions/withdraw/
post_comment Publish a (public or private) comment, e.g. a bot's forecast rationale; supports replies and attaching your latest forecast. POST /api/comments/create/
get_comments Read comments, filtered per the API's access rule (author=<your id> and/or author_is_staff=true). GET /api/comments/
download_question_data Fetch the CSV Zip export (question/forecast/score data), unzip it in the Worker, and return parsed rows. Restricted endpoint — see caveats. GET /api/data/download/

The MCP handshake and tools/list work without a token (so registries can scan the server); every tool call requires one and returns setup instructions if it's missing.

Authentication

  1. Create a free Metaculus account, then generate an API token at https://www.metaculus.com/accounts/settings/account/#api-access (URL from the official API spec).
  2. Configure your MCP client to send the token on every request. Accepted headers (first match wins):
    • Authorization: Token <token> — Metaculus's native scheme
    • Authorization: Bearer <token>
    • X-Metaculus-Token: <token>

The entire Metaculus API rejects unauthenticated requests (verified live 2026-07-12: GET https://www.metaculus.com/api/posts/?limit=1 without a token → HTTP 403 "The API is only available to authenticated users.").

Connect from Claude Code

claude mcp add --transport http metaculus \
  https://metaculus-mcp.anishboddu6.workers.dev/mcp \
  --header "Authorization: Token YOUR_METACULUS_TOKEN"

Or in .mcp.json:

{
  "mcpServers": {
    "metaculus": {
      "type": "http",
      "url": "https://metaculus-mcp.anishboddu6.workers.dev/mcp",
      "headers": { "Authorization": "Token ${METACULUS_TOKEN}" }
    }
  }
}

(Syntax verified against the Claude Code MCP docs, code.claude.com/docs/en/mcp, accessed 2026-07-12.)

Deploy

npm install
npm run typecheck        # tsc --noEmit
npm run build            # wrangler deploy --dry-run --outdir=dist
npx wrangler deploy      # needs a wrangler-authenticated machine

No bindings, no secrets, no Durable Objects — a single stateless Worker. observability.enabled = true gives you usage logs in the Cloudflare dashboard (tool calls log only method + whether a token was present, never token values).

Test locally

npx wrangler dev --port 8799
# 1) MCP handshake
curl -s -X POST http://localhost:8799/mcp \
  -H "Content-Type: application/json" -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"t","version":"1"}}}'

# 2) List the 8 tools
curl -s -X POST http://localhost:8799/mcp \
  -H "Content-Type: application/json" -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}'

# 3) Real call (uses YOUR token against the live API)
curl -s -X POST http://localhost:8799/mcp \
  -H "Content-Type: application/json" -H "Accept: application/json, text/event-stream" \
  -H "Authorization: Token YOUR_METACULUS_TOKEN" \
  -d '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"list_questions","arguments":{"statuses":["open"],"limit":3,"order_by":"-hotness"}}}'

Honest caveats (all from the official API spec, accessed 2026-07-12)

  • Terms of Use govern the data. Per the spec's own warning: API data may not be used to train, evaluate, or otherwise create or develop AI/ML models without Metaculus's prior written permission, and any commercial use requires a separate written agreement (https://www.metaculus.com/terms-of-use/). Using questions as live LLM context to forecast is exactly what Metaculus's own bot tournaments invite — but do not resell or train on the data.
  • Community Prediction is tier-limited. Normal authenticated accounts see CP aggregates on only ~50 questions. A free "Bot Benchmarking Access Tier" (requested via the Metaculus Data Needs Google Form linked at metaculus.com/api/) extends CP/resolution access to ~250 open + ~250 resolved questions. Your own predictions/scores/comments are always available to you.
  • download_question_data (and the email variant) are restricted endpoints — the spec directs users to the same Data Needs Form for access, and project_id exports additionally need per-project whitelisting. Expect 4xx errors without that grant.
  • Rate limits: UNKNOWN. The spec states request throttling applies to metaculus.com but publishes no numbers. The server surfaces 429s verbatim with a back-off hint.
  • No free-text search. The documented GET /api/posts/ has structured filters only; there is no documented search parameter.
  • Comments access rule. When API access is restricted, GET /api/comments/ requests must filter by author=<your own user id> and/or author_is_staff=true.

Publishing to registries

Current state: this server has not been submitted anywhere yet. All three processes below were verified 2026-07-12; each requires human steps (accounts/logins).

  1. Official MCP Registry (registry.modelcontextprotocol.io) — publish server.json (template included in this repo; fill in your GitHub username + workers.dev URL) with the mcp-publisher CLI. GitHub-based authentication requires the server name to start with io.github.<your-username>/. Remote-only servers need no npm package — the remotes entry with your public Workers URL is enough, and the included headers entry (isRequired/isSecret) tells clients to collect the user's token. Source: modelcontextprotocol/registry repo docs (quickstart.mdx, remote-servers.mdx).
  2. mcp.so — submissions "only support public GitHub MCP servers": push this directory to a public GitHub repo, sign in at mcp.so, submit the repo URL at https://mcp.so/submit, then complete the draft (saving publishes it). Source: the mcp.so submit page, fetched 2026-07-12.
  3. Smithery (smithery.ai) — go to https://smithery.ai/new, enter the server's public HTTPS URL, and complete the publishing flow (requirement: Streamable HTTP transport — this server complies). Smithery scans the server to extract tool metadata; because this server answers initialize/tools/list without a token, the automatic scan can complete. Source: smithery.ai/docs/build/publish.md, fetched 2026-07-12.

So the one hard human prerequisite across registries: a public GitHub repo (required by mcp.so; also the natural namespace proof for the official registry) and the respective account logins.

Monetization (honest note)

MCP has no native billing layer. v1 is free and usage-logged (Cloudflare observability shows request counts and authed/unauthed split). If usage materializes, later options include a hosted "pro" tier (e.g. caching, scheduled scans, portfolio alerts) — but note the Metaculus ToS clause above: any commercial use of Metaculus API data itself requires a written agreement with Metaculus, so monetize the tooling/infrastructure only after securing that, or keep it free as a distribution/reputation asset.

Architecture

  • src/index.ts — Worker entry: token extraction from headers, / info page, /mcp endpoint via createMcpHandler (Cloudflare Agents SDK, Streamable HTTP transport). A fresh McpServer per request (MCP SDK ≥ 1.26.0 requirement for stateless servers).
  • src/server.ts — the 8 tool registrations (zod schemas + response slimming to keep agent context small).
  • src/metaculus.ts — thin authenticated fetch client for https://www.metaculus.com/api.
  • src/cdf.ts — TypeScript port of the CDF algorithms documented in the official API spec: nominal→scaled location (linear + logarithmic zero_point scaling), percentile-sketch interpolation, the spec's bound/slope standardization step, and validation of all documented CDF rules (16 unit checks pass, including the spec's own scaling examples).
  • Stack: agents 0.17.3, @modelcontextprotocol/sdk 1.29.0, zod 4, fflate (CSV-zip parsing), TypeScript 5 strict, wrangler v4.

Sources (all accessed 2026-07-12)

  • Official Metaculus OpenAPI 3.0.0 spec (version 2.0.0): https://raw.githubusercontent.com/Metaculus/metaculus/main/docs/openapi.yml (rendered at https://www.metaculus.com/api/) — endpoints, parameters, payload schemas, TokenAuth scheme, access tiers, ToS restrictions, CDF algorithm and rules, restricted-endpoint notes.
  • Live auth behavior: GET https://www.metaculus.com/api/posts/?limit=1 → HTTP 403 (unauthenticated).
  • Cloudflare Agents SDK MCP docs: https://developers.cloudflare.com/agents/model-context-protocol/apis/handler-api/ (createMcpHandler, stateless pattern, SDK 1.26.0 breaking change) and .../apis/agent-api/.
  • MCP Registry publishing: https://github.com/modelcontextprotocol/registry — docs/modelcontextprotocol-io/quickstart.mdx and remote-servers.mdx.
  • mcp.so submission requirements: https://mcp.so/submit.
  • Smithery publishing: https://smithery.ai/docs/build/publish.md.
  • Claude Code client config syntax: https://code.claude.com/docs/en/mcp.md.

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选