Documentation packages for AI agents.
The fastest way for agents to read the docs.
mandex distributes library documentation like packages — versioned, cached locally, searchable offline in 40ms. No API keys, no rate limits, no network dependency.
Your AI agent gets the right docs for the right version, instantly.
curl -fsSL https://mandex.dev/install.sh | shOr with Cargo:
cargo install mandex# Pull docs for a library
mx pull fastapi
# Search across installed docs
mx search fastapi "rate limiting"
# Show a specific entry
mx show fastapi "Rate Limiting"
# Auto-sync docs for all project dependencies
mx sync| Cloud docs (MCP/API) | mandex | |
|---|---|---|
| Latency | 300–500ms per query | 40ms (local FTS5 + neural reranking) |
| Offline | ❌ | ✅ |
| Rate limits | Per-provider quotas | Unlimited |
| Versioning | Latest only or manual | Pinned to exact version |
| Setup | API keys, tokens, config | mx pull <package> |
- Pull — download a compressed documentation index (SQLite + FTS5) from the registry
- Search — full-text search with BM25 ranking, automatically reranked by a local neural cross-encoder for semantic relevance
- Show — retrieve the full documentation entry for an exact match
Documentation is stored as local SQLite databases in ~/.mandex/cache/. A per-project .mandex/index.db merges all project dependencies into a single index for fastest search.
mandex works with any AI coding assistant that can run shell commands:
# Set up integrations (Claude Code, Cursor, Windsurf, Codex)
mx init
# Or just use mx directly in your agent's tool calls
mx search nextjs "server actions"mx init installs a skill for Claude Code, cursor rules for Cursor, and agent instructions for Codex — so your agent automatically searches mandex before writing code.
| Command | Description |
|---|---|
mx search <package> "<query>" |
Search within a package |
mx search "<query>" |
Search across all installed packages |
mx show <package> "<entry>" |
Show full entry content |
mx pull <package>[@version] |
Install docs for a package |
mx sync |
Auto-detect and install docs for project dependencies |
mx list |
Show installed packages |
mx info <package> |
Show package details |
mx remove <package> |
Remove a package |
mx init |
Set up AI assistant integrations |
mx build <dir> --name <n> --version <v> |
Build a .mandex package from markdown |
55 packages across npm, pip, and cargo — browse the full registry.
View all packages
npm: ai-sdk, angular, astro, better-auth, claude-code, clerk, drizzle-orm, express, fumadocs, hono, langchain-js, mongodb, nest, nextjs, nuxt, openclaw, opencode, playwright, prisma, react, shadcn-ui, supabase, svelte, tailwindcss, tanstack-query, trpc, turborepo, vite, vitest, vue, zod, zustand
pip: celery, django, fastapi, flask, httpx, instructor, langchain, langgraph, llama-index, numpy, pandas, pydantic, pytest, requests, scipy, sqlalchemy, streamlit, transformers, uvicorn
cargo: axum, mandex, serde, tokio
- 40ms average search latency (local FTS5 + reranking)
- 70ms project-wide search across all dependencies (merged index)
- 5ms tokenizer load (tokie
.tkzbinary format) - Pre-optimized ONNX reranker with mmap + multi-threaded inference
Have documentation for a library? Build and publish it:
# Build from a directory of markdown files
mx build ./docs --name my-lib --version 1.0.0
# Output: my-lib@1.0.0.mandex (compressed SQLite + FTS5 index)mandex accepts any directory of .md / .mdx files — Docusaurus, MkDocs, Mintlify, Starlight, or plain markdown all work.
Contributions welcome. See the issues for open tasks.
MIT