MCP server for fact-checking AI claims against authoritative sources. Works with any MCP-compatible agent — Claude, Cursor, Continue, Cody, Zed, Windsurf, OpenAI Agents, LangChain, and more.
FAQ (answers come from our records)
How do I install it?
The repository provides this install command: claude mcp add webcite -- npx -y webcite-mcp-server. It comes from the repository itself; we cannot guarantee it runs as-is in your client.
Who is it for?
We filed it under these tags: 开源项目. The tags come from the entry itself, not a taxonomy we imposed.
Who maintains it?
repository strategyconnect/webcite-mcp-server;first catalogued 2026-10-10;last seen 2026-10-10。There is no official certification or quality rating here — we record sources and dates only.
WebCite MCP Server
以下摘录自该仓库的 README(片段) · 在 GitHub 读完整 README →
> Remote Streamable HTTP MCP: **https://api.webcite.co/mcp**. Start at **https://webcite.co/connect**. The remote server uses the public profile for all supported public API workflows. Check the published npm version before pinning it, because repository changes can precede publication.
MCP (Model Context Protocol) server for WebCite — lets any AI agent verify factual claims against authoritative sources, bind quotes back to the passage they came from, and read the numbers out of documents deterministically.
Works with **any MCP-compatible client** including Claude Desktop, Claude Code, Cursor, Continue, Cody, Zed, Windsurf, OpenAI Agents SDK, LangChain, and more.
Verification and document acceptance
On servers implementing retained generation navigation, a folder generation can include one navigation-query provider stage within the existing five-credit generation operation. Its route is selected before calling the provider, and a refusal, invalid reply, failure or unknown outcome remains an error without provider switching, automatic retry or silent original-query fallback. Public memo model and token usage keep their existing meaning; separate query-provider measurements are internal. Internal staged continuation is identity-bound and requires complete accounting with no draft dispatch. It does not expose a new public preparation feature. Confirm the serving release before relying on this behavior.
On servers implementing isolated generation retrieval budgets, the default 16-anchor retrieval limit gives evidence-reference hints and mandatory core context up to 32,000 retained-text characters, plus up to 32,000 incremental ordinary-retrieval characters, bounded by 64,000 in total. Without evidence references, retained text remains bounded by 32,000 characters. Ordinary retrieval allows up to 16 anchors per retained navigation query, or 16 with no navigation queries, subject to the same text and physical-unit ceilings. The physical-unit ceiling is 48 plus three times the reference count; each passage is capped at 2,000 UTF-16 characters. Stored source inventory retains its separate 100 MiB aggregate read guard. These are bounded retrieval allowances, not guarantees of exhaustive source coverage, complete prompt size or provider-token cost. Confirm the serving release before relying on them.
For an official-source requirement, supply filters on verify_claim or verify_claim_stream, for example {"official_country":"ae","is_primary_source":true,"domain":["centralbank.ae"]}. Pass a known public HTTPS report in source_urls; writing a URL or site: expression in the claim is not a substitute for these fields. Match the quoted metric, observation period, geography, currency and scale. Different official publications can remain unresolved rather than yielding one universal corrected figure.
Upload original image, PDF or PowerPoint bytes and wait for parsing before review. Inspect input coverage, unreadable regions, claim limits and per-claim source/page locators. A successful upload, a completed review, or many citations does not certify that every figure was read or verified. Citation links are not a guarantee of pixel-level highlighting on an uploaded slide. An extraction failure stays a failure; a bounded search with no evidence does not prove the data is unavailable. Open the original publisher passage before treating a stance badge as proof.
Preserve saved review, citation, job and operation IDs on a timeout or uncertain outcome. Read retained progress before retrying and reuse an explicit idempotency key only with the same request. Package publication does not deploy the hosted API, frontend or MCP binary. Check their actual serving identities before asserting that reported cases are fixed in production.
1.9.57 release notes
Updates the guide for original-document acceptance and scoped application access. App-bound keys cannot use legacy account administration or share personal UUID citations. Folder traces respect retained source and linked-memory visibility. Confirm the serving release before relying on these controls. Tool names and client interfaces are unchanged.
1.9.56 release notes
Adds hosted MCP upload_url and publishes the reviewed timeout and evidence workflows:
- upload_url stores bounded original bytes from public HTTPS URLs through public-address and redirect checks, returning owned asset references and usage. Upload success does not certify parsing or a full review.