How to Install the Fetch MCP Server

Let your AI assistant retrieve and convert web pages with the official Fetch MCP server.

OVERVIEW The Fetch MCP Server is an official Python reference server in modelcontextprotocol/servers. The main README describes it as "Web content fetching and conversion for efficient LLM usage" — it retrieves HTTP(S) URLs and converts HTML into text suitable for model context. Commands in this tutorial are verified from https://github.com/modelcontextprotocol/servers only. WHAT IT DOES Fetch provides a focused tool surface (documented as a single primary fetch tool in directory metadata) for pulling remote documentation, blog posts, release notes, and articles into MCP workflows. Instead of pasting large web pages into chat, the assistant invokes Fetch with a URL and receives cleaned text. Fetch does not replace a search engine — you supply URLs. For search integration, browse community servers on the MCP Registry; the archived Brave Search reference server moved to servers-archived. PREREQUISITES - uv / uvx (recommended in official README) OR pip - Network access from the machine running the server - MCP client configured to launch the server Install uv: https://docs.astral.sh/uv/getting-started/installation/ Verify: uvx --version STANDALONE INSTALL COMMAND (OFFICIAL) Python servers in the official README use uvx: uvx mcp-server-fetch Alternative pip installation is documented generically for Python servers in the same README (pip install + python -m module pattern). For Fetch, uvx is the documented one-liner in the reference server list. CLAUDE DESKTOP CONFIGURATION { "mcpServers": { "fetch": { "command": "uvx", "args": ["mcp-server-fetch"] } } } No path arguments required. Add alongside other servers under mcpServers. WINDOWS NOTE uvx does not use the cmd /c wrapper required for npx on Windows. Only npx-based TypeScript servers need cmd /c per the official README. CURSOR NOTES Add Fetch to mcp.json for documentation-heavy coding tasks. The agent can fetch vendor docs before suggesting APIs. Consider network proxy environment variables if your corporate network requires them — set env in MCP config if your client supports it (pattern shown for other servers in README with env blocks). CLINE NOTES Cline agents researching bugs can Fetch issue tracker pages or stack overflow threads (respect terms of service). Enable Fetch when you want autonomous URL retrieval; disable if you prefer manual paste-only workflows. SECURITY NOTES Fetching URLs chosen by an LLM introduces SSRF risk — internal IP addresses might be reachable from your laptop. Block sensitive network ranges at the firewall if deploying Fetch in corporate environments. Reference servers carry the official disclaimer: evaluate threat models before production use. Do not embed credentials in URLs passed to Fetch. FAQ Q: Does Fetch execute JavaScript? A: No — static HTML conversion. Dynamic SPAs may return incomplete content. Q: Official install command? A: uvx mcp-server-fetch from https://github.com/modelcontextprotocol/servers README. Q: Fetch vs Puppeteer MCP? A: Puppeteer reference server is archived in servers-archived for browser automation. Fetch is lightweight HTTP fetch. TROUBLESHOOTING Test uvx mcp-server-fetch in a terminal. First launch downloads Python dependencies. If Claude shows no tools, verify JSON config and restart the client completely. WORKFLOW TIPS Pair Fetch with Memory to remember key facts from docs, and Sequential Thinking (npx -y @modelcontextprotocol/server-sequential-thinking) for multi-document comparisons. HOW FETCH FITS THE REFERENCE SERVER SET Among the seven active reference servers listed in the main README — Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking, Time — Fetch is the only one focused on outbound HTTP content retrieval. That makes it the default official choice for documentation fetching tasks until you adopt a community search MCP. CONTENT CONVERSION BEHAVIOR Fetch converts HTML into LLM-friendly text so models spend tokens on substance rather than markup. Complex pages with heavy navigation chrome may still include boilerplate — follow up by asking the model to summarize relevant sections. For authenticated documentation portals, Fetch cannot access content behind login walls unless you provide session mechanisms outside MCP scope (generally discouraged). RATE LIMITING AND POLITENESS Respect robots.txt and vendor rate limits when agents loop Fetch across many URLs. Aggressive autonomous fetching from CI agents can trigger blocks — throttle tool approval in Cline or Cursor Agent when researching large doc sets. EXTENDED FAQ Q: Can Fetch POST forms? A: The reference Fetch server targets retrieval and conversion — not general HTTP clients. Q: pip instead of uvx? A: README documents pip patterns for Python servers generically; uvx mcp-server-fetch is the documented one-liner. Q: Fetch on air-gapped machines? A: Requires network egress to target URLs at minimum; initial uvx install also needs package download unless cached. COMPARISON WITH ARCHIVED BRAVE SEARCH Brave Search lived in servers-archived as a search API MCP. Fetch plus a search-capable community server reproduces "discover then read" pipelines. For known URLs, Fetch alone suffices. OFFICIAL README CITATION The exact install line uvx mcp-server-fetch appears in the Getting Started and Reference Servers sections of https://github.com/modelcontextprotocol/servers. Do not substitute alternate PyPI package names without verifying they are the same mcp-server-fetch distribution maintained in that monorepo. TEAM POLICY TEMPLATE Document allowed URL domains for autonomous Fetch in agent workflows. Block file:// and internal RFC1918 ranges at the network layer where possible. Log Fetch tool arguments in client audit trails when compliance requires provenance for external content used in generated code.