Fetch

The Fetch server is one of the seven active reference servers. It lives at src/fetch in the modelcontextprotocol/servers monorepo and is distributed as mcp-server-fetch on PyPI. The common pairing is "search MCP + Fetch" — the search server returns URLs, Fetch turns each into clean markdown the agent can actually reason over. What it does: - Fetches an arbitrary URL and returns readability-style markdown, stripping nav and ads - Consults robots.txt before each request and refuses disallowed paths - Caps response size at 1 MiB by default, configurable per request - Exposes a raw mode that returns the original HTML when the markdown extractor would lose fidelity - Follows up to 5 redirects, surfaces the final URL for citation - Sends a realistic User-Agent so sites that block default Python UAs still work Example: Pair with the Brave Search entry in this registry. The agent runs a web search, picks the top three results, and calls fetch on each to extract the body text — it then summarises across the three articles instead of guessing from snippet text. Setup: No required configuration. Optional FETCH_TIMEOUT (seconds) and FETCH_MAX_SIZE (bytes) env vars adjust defaults. Run via "uvx mcp-server-fetch" for an isolated-env install or "pip install mcp-server-fetch" into an existing Python environment. Transport: stdio only. Single-process, no background workers. Notes: The server does not render JavaScript. For client-rendered pages that need a real browser (SPAs, dashboards behind auth, sites that ship skeleton HTML) use the Puppeteer entry in this registry instead. The robots.txt check is strict — sites that disallow your user-agent will return an explicit "robots disallowed" error rather than silently returning empty content.