DocsForgeDocumentation for technologies your model was never trained on.
DocsForge finds a library's official docs from its name alone, verifies the page really documents it, harvests the whole thing, and hands it to any MCP client — Claude Code, Cursor, Codex and other AI coding assistants — with a coverage figure you can gate on.
learn_technologya recorded runHow it works
Name a library. DocsForge works out where it documents itself, confirms the page really documents it, and harvests the whole set.
01Resolve
Asks the package registries — npm, PyPI, crates.io, all keyless — where the library declares its documentation, then probes the site for an llms.txt or a docs root.
02Verify
A page must name the package repeatedly before it is trusted. If nothing verifies, DocsForge says so and asks for a URL instead of picking something plausible.
03Harvest
Best strategy first: llms-full.txt, then a docs-scoped sitemap, then a scoped crawl. Every page lands as Markdown, versions kept apart, incompleteness reported — never hidden.
Every source type, detected automatically
The URL and, when needed, the content decide the strategy. You never pick.
Honest about coverage
A third of a manual that looks whole produces confident, wrong answers. So a partial harvest says so — in the harvest result, in list_knowledge_base, and again every time the content is read.
Versions are stored side by side. v2 and v3 contradict each other, and a model handed both will quote the wrong one.
One engine, every surface
One extraction engine backs every surface, so an MCP client and the chat panel return byte-identical results. Several DocsForge instances can share one store.
Frequently asked questions
What is DocsForge?
DocsForge is an open-source documentation MCP server. Give it the name of a library, framework or tool and it finds the official documentation through the package registries, verifies that the page really documents that library, harvests the whole set as Markdown, and serves it to your AI coding assistant through MCP tools.
Who is it for?
Developers whose AI coding assistant writes code against libraries newer than its training data, or mixes up their versions; AI engineers building agents that must answer from real documentation rather than from memory; and teams that want one shared, searchable store of the docs their projects depend on.
What is an MCP server, and which tools does DocsForge provide?
The Model Context Protocol (MCP) is an open standard that lets an AI assistant call tools on a server. DocsForge provides thirteen: learn_technology learns a library from its name alone, search_knowledge_base and read_knowledge_base answer from what is stored, scan_project reports a project's dependencies and which are already documented, and the rest fetch, harvest and manage documentation. They are served over Streamable HTTP or stdio.
Which AI coding assistants does it work with?
Any MCP client. The connect page has a ready command or config block for Claude Code, Codex, ChatGPT, Cursor, Windsurf, Antigravity, Gemini CLI, VS Code and Claude Desktop.
How does it avoid handing the model the wrong documentation?
A page must name the package repeatedly before it is trusted; if nothing verifies, DocsForge says so and asks for a URL instead of guessing. Versions are stored apart, so v2 and v3 are never mixed, and a harvest that stopped early is marked incomplete every time it is read.
Does it need an API key?
Not to resolve a name: npm, PyPI and crates.io are asked without a key. A hosted DocsForge asks clients for a bearer token; running it yourself on your own machine needs none.
Is DocsForge free and open source?
Yes. DocsForge is MIT-licensed. Install it with pipx install docsforge, run it on your own machine or host it yourself, and keep harvested documentation as Markdown files or in PostgreSQL.
Built by Raj Patel
DocsForge is designed, built and maintained by Raj Patel. Found a bug or a bad extraction, have an idea, or want to contribute? Open an issue on GitHub, or reach out directly. Meet everyone who has contributed.
Connect your MCP client.
Streamable HTTP. One URL, one bearer token. The same thirteen tools whichever client you use.