Auto-Refreshing Library Documentation MCP Server Because Training Data Is Stale and Stack Overflow Is Dead
AI coding agents (Claude Code, Cursor, Copilot) keep generating plausible-but-wrong code that calls removed APIs, uses deprecated parameters, or invents syntax. Core reason: their training data is months-to-years old, and Stack Overflow's decline means there's no fresh human-written corrective signal. Builders are scrambling to fill the gap — Context7, Instagit, Ref Tools each attack a slice — but coverage is fragmented and each supports a different subset of ecosystems. The universal version: a single MCP server that auto-pulls latest docs for every npm/PyPI/crates.io/Go module, version-resolves to the user's lockfile, and serves fresh documentation to any agent.
Don't try to index the whole internet. Index docs of packages on PyPI / npm / crates.io / Go proxy, keyed by version. When an agent asks, parse the user's lockfile first, return THAT version's docs. That alone eats 80% of 'agent hallucinated a removed API' failures. The moat is the fetch+parse pipeline for 20+ docs site formats, not the MCP wrapper.
landscape (5 existing solutions)
Everyone agrees the problem is real — stale training data produces broken code. The category exploded in Q1 2026 but every entrant attacks one ecosystem. The 'universal' version (one MCP server, resolves to your lockfile, pulls fresh from every package registry) is the consolidation play nobody has landed yet. Harder than it sounds because package-level docs are in a dozen different formats (README, docs sites, Sphinx, mkdocs, TSDoc, rustdoc).