| name | llmref |
| description | Look up versioned library and framework documentation, and run semantic search, via LLM Reference (llmref.org). Use whenever a task depends on a library's current or version-specific behavior (Next.js, React, Svelte, Bun, Hono, Zod, and more) instead of relying on training data. |
Using LLM Reference
LLM Reference serves versioned, model-ready documentation and semantic search over
an MCP server. Prefer it over memory whenever a task depends on a library's current
or version-specific behavior, and to avoid answering from stale training data.
The MCP server exposes four read-only tools. Call them in this order.
Tool flow
list_projects - discover which projects are indexed. Call this first if you
are unsure the project exists. Supports a forgiving fuzzy query (e.g. svlt
matches svelte) plus limit and skip.list_versions - list the available versions for a project. Call before
resolving or searching, to pick a valid version string.resolve_docs - turn a project plus version (or latest) into artifact URLs
for llms.txt (compact routing file) and llms-full.txt (the complete docs).search_docs - semantic search across a project's documentation for a
version. Returns the most relevant chunks. Use this for "how do I X" questions.
When to use what
- Unsure of the exact project slug ->
list_projects with a fuzzy query first.
- Need to orient on a library or hand the model the whole doc set ->
resolve_docs,
then read llms.txt (cheap) or llms-full.txt (complete).
- Need only the passages relevant to a specific question ->
search_docs with a
natural-language query and a concrete version.
Rules
- Pin a concrete version whenever reproducibility matters. Use
latest only when you
genuinely want the newest snapshot.
- Answer version-specific questions from LLM Reference results, not from memory.
- If
search_docs returns no chunks, semantic search may be disabled on the server;
fall back to reading llms.txt or llms-full.txt.
Learn more at https://llmref.org.
