Provides optional experimental content extraction through a configured Ollama instance, enabling AI-powered extraction of web pages via the web_extract tool.
Integrates with a configured SearXNG instance to provide web search results through the web_search tool, while local offline full-text search remains available.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Fetchkeep fetch https://example.com/article and save a clean markdown copy"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
The web, saved for your agents.
Fetchkeep is an open-source, local-first web retrieval tool for AI agents. It fetches pages, extracts clean Markdown, saves every version to a local SQLite store, searches it offline, crawls sites within strict bounds and returns citations you can re-read later — through a CLI and an MCP server.
No browser, LLM or paid service required. HTTP fetching, extraction, saving, reading, search and crawling work out of the box. Browsers (Chromium, experimental Lightpanda) and Ollama are optional add-ons.
Stable citations. Every document version is content-addressed; every block has a reference
(fk:d_…@2#b7) that returns exactly the same text tomorrow, offline.
Safe by default. Private and internal addresses are blocked on every redirect and resolved address, sizes and time are bounded, fetched content is fenced as untrusted data.
Honest results. One response envelope for CLI and MCP with status (success / partial / error),
backend attempts, timings, truncation and actionable errors.
Status: v0.1.1. Apache-2.0.
Requires Node.js ≥ 22.19 (current LTS lines 22 and 24). No native build step: storage uses Node's built-in
node:sqlite.
npm install -g fetchkeep
fetchkeep doctorThe package is available on the npm registry. From a checkout: npm ci && npm run build && npm link.
The install pulls no browser and requires no compiler.
Related MCP server: Scout MCP Server
fetchkeep fetch https://en.wikipedia.org/wiki/Markdown > markdown.md # Markdown on stdout, metadata on stderr
fetchkeep search "lightweight markup language" # offline full-text search
fetchkeep read https://en.wikipedia.org/wiki/Markdown --find "John Gruber" # exact quote + block reference
fetchkeep read "fk:d_…@1#b4" # re-read a cited block
fetchkeep crawl https://docs.example.org/ --max-pages 30 --max-depth 2
fetchkeep --json fetch https://example.com # full JSON envelopeOther commands: list, versions, export (JSON Lines), delete, prune, crawls, extract (Ollama,
experimental), doctor, mcp. Exit codes: 0 success, 1 error, 2 usage, 3 partial. Workspaces isolate stores:
fetchkeep -w research fetch ….
{
"mcpServers": {
"fetchkeep": { "command": "fetchkeep", "args": ["mcp"] }
}
}Tools: web_fetch, web_read, web_search, web_crawl, web_extract (experimental). Each returns the envelope as
structuredContent plus a text rendering in which page content sits inside <untrusted-web-content> delimiters.
stdout carries only protocol messages; logs go to stderr. See docs/mcp.md.
auto (default) fetches over HTTP and escalates to an enabled browser only when the page looks unrendered
(empty SPA root, "enable JavaScript" notices, very little text on a scripted page, or HTTP 403/429/503). Without a
browser it returns the HTTP result as partial and says why.
npm install -g playwright-core
npx playwright-core install --only-shell chromium # headless shell only
export FETCHKEEP_CHROMIUM=1Lightpanda is supported experimentally through CDP (FETCHKEEP_LIGHTPANDA_EXECUTABLE or
FETCHKEEP_LIGHTPANDA_ENDPOINT, plus puppeteer-core). Details, isolation guarantees and the escalation heuristic:
docs/browsers.md.
bench/ compares Fetchkeep's profiles with Firecrawl (self-hosted and hosted, separately) and DonSeTch on
deterministic local fixtures and an annotated set of live websites. See bench/README.md and the
published results in bench/results/.
npm run bench -- --engines fetchkeep,firecrawl,donsetch --suite fetch
npm run bench:report -- --run bench/runs/<run>Platform | Core (HTTP, store, search, crawl, MCP) | Chromium backend | Lightpanda backend |
Linux x64 | tested (CI: Ubuntu) | tested (Ubuntu 24.04) | tested (experimental) |
Linux arm64 | untested | untested | untested |
Windows 10/11 x64 | tested (CI: Windows) | tested (Windows 11) | via WSL2 only |
macOS | expected to work (untested) | expected to work (untested) | Lightpanda publishes macOS builds (untested) |
JavaScript rendering needs an optional browser; HTTP mode does not execute scripts by design.
No OCR: scanned PDFs yield no text. No login flows, forms or page interactions.
Web search is only available through a SearXNG instance you configure; local search is always available.
Lightpanda support is experimental and unsandboxed.
See CONTRIBUTING.md and SECURITY.md.
Apache License 2.0 — see LICENSE and NOTICE. Third-party licenses: THIRD_PARTY_LICENSES.md.
This server cannot be deployed
Scrape, crawl and search the web for AI agents via MCP.
Turn any public website into an MCP server for agents to search, read and navigate.
Web data tools for AI agents: pages as markdown, search, maps, commerce, jobs, AI answers.
Fetch pages as markdown, search web and news, extract structured data. For AI agents.
