web-researcher-mcp
A research MCP server: searches the web, SEC, courts and patents, reads whole pages and checks that cited sources really exist.
- 77 ★ on the repository
- Last commit
Broad access
Why this level
- Queries go to the chosen search services, and with `site_login` encrypted cookies of your accounts are kept on disk.
- An agent with your login reads LinkedIn, X and other pages on your behalf.
Install
- Works in
Manual install
claude mcp add --scope user web-researcher -- uvx web-researcher-mcpRequires `uv` installed, no keys needed.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Queries go to the chosen search services, and with `site_login` encrypted cookies of your accounts are kept on disk.
- An agent with your login reads LinkedIn, X and other pages on your behalf.
This is third-party code. Review the repository files before installing.
What it does
web-researcher-mcp is an MCP server that answers with verifiable sources. It searches the web and primary databases: SEC filings, US court decisions, patents, clinical trials and full-text papers. Search can be limited to sites you trust, and when one search engine fails the next takes over. The server reads whole pages: PDF, Word, PowerPoint, YouTube transcripts and Hacker News threads. The `verify_citation` and `audit_bibliography` tools check that a source exists, was not retracted and has a live link. Optionally `site_login` saves your own login to LinkedIn, X and other sites so the agent reads pages you can already open. Research sessions survive context loss and citations export to BibTeX, APA, MLA, RIS and CSL-JSON.
Who it is for. Researchers, analysts and lawyers who need AI answers with sources that can each be opened and verified.
Good fit when
- A literature review is being prepared and links must lead to real DOIs, not invented ones.
- A client report needs sources that open with a click.
- A SEC document or a US court case has to be found by query.
- A BibTeX reference list is audited as a whole for retracted and dead links.
- Answers of several models are compared with `research_panel`.
Not a fit when
- You only need a list of search links: a plain search API is enough.
- Social network logins cannot be stored even encrypted on your machine.
- Findings go to publication without reading sources: the server returns evidence, not a verdict.
Example request
Find primary sources on CRISPR off-target effects and for each link check with `verify_citation` that the paper exists and was not retracted.MCP
- Transport
- stdio, http
- Authentication
- not required
Environment variables
SEARCH_PROVIDER- Search provider: brave, serper, tavily, exa and others. DuckDuckGo works when unset.
BRAVE_API_KEYsecret- Key of the chosen provider, the variable name depends on it.
COOKIE_VAULT_ENABLED- Enables `site_login` and the login vault.
Limitations
Without keys DuckDuckGo search works, but other engines need their own keys, for example `BRAVE_API_KEY` or `TAVILY_API_KEY`. The `citation_graph` and `research_panel` tools need at least one configured provider, `site_login` turns on only with `COOKIE_VAULT_ENABLED=true`, and memory with `MEMORY_ENABLED=true`. With `site_login` cookies or a browser profile are saved encrypted, so enable it only on your own machine. Citation checks return evidence and the final judgment stays with the reader.
How to disable. Remove the MCP entry with `claude mcp remove web-researcher` and stop any shared server you ran. Delete saved logins with the `site_logout` tool and clean cache and memory by hand.
FAQ
How do I connect web-researcher-mcp to Claude Code?
Run `claude mcp add --scope user web-researcher -- uvx web-researcher-mcp`. It needs `uv` and the server downloads the rest. Other clients use the same `uvx web-researcher-mcp` command.
Is an API key required?
No, DuckDuckGo search works without keys. For Brave, Serper, Tavily, Exa and other engines set `SEARCH_PROVIDER` and a key, for example `BRAVE_API_KEY`.
How do I read pages behind a login?
Set `COOKIE_VAULT_ENABLED=true` and call `site_login`: you sign in yourself, 2FA included, in your own browser or a private window. The password is not saved. Cookies or the profile are stored encrypted.
What does verify_citation check?
It checks that the source exists, was not retracted, the link is live and the text says what is claimed. The result is evidence, not a final verdict.
How do I continue research after context loss?
Call `get_research_session` and the session resumes where it stopped. `research_export` exports the result as markdown or JSON with provenance for every step.
Discussion
No questions or reviews yet.
Related
MCP serversA local MCP server with no API keys: multi-engine web search, page fetching and document reading tuned for LLMs.
MCP server with DuckDuckGo search and page text extraction: several URLs can be passed in one call
A unified search MCP server: one interface to Tavily, Brave, Kagi, Exa, Linkup, Firecrawl and GitHub search
An MCP server that searches Google and Scholar, reads pages and PDFs and stores findings in a local project knowledge graph
Linkup MCP server: real-time web search, page fetching and long research tasks with sources for an agent
A web search MCP server that gives the agent full page text at once: Stack Overflow answers, GitHub threads, arXiv papers
README in short
The server ships as a single signed binary through `uvx`, Homebrew, install scripts or Python wheels. It has a `WebResearcherClient` Python client, an optional hash-chained call log and 37 tools.