memo

MCP server lokal untuk dokumentasi library versi-terbaru — klon Context7 yang gratis total, unlimited, dan offline-capable. Index dibangun on-demand ke SQLite (BM25 + vektor) dari llms.txt / halaman docs resmi.

Fitur

• resolve_library_id(name, query) — resolusi library → kandidat {repo, docs_url, trust}: alias kurasi manual → builtin stdlib (Node/Python) → directory.llmstxt.cloud → npm/PyPI (trust = downloads) → GitHub search (preferensi bahasa dari query)
• get_docs(library_id, query) — hybrid search (BM25 + embedding bge-small-en-v1.5), cache sub-ms setelah first fetch
• versions(library_id) — daftar versi yang diketahui
• Offline-capable: sekali di-index, cache permanen di SQLite (FTS5 + sqlite-vec)

Instalasi

uv tool install git+https://github.com/ngabzar02/memo-server

Registrasi di opencode.json (atau config MCP lain):

{
  "mcp": {
    "memo": {
      "type": "local",
      "command": ["<path ke uv tool bin>/memo"],
      "enabled": true
    }
  }
}

Pre-index (disarankan)

Cold fetch pertama melewati batas timeout MCP (~30s) di beberapa client — panaskan cache dari shell dulu:

memo --warmup flask nextjs httpx
memo --warmup --force flask   # re-ingest (docs_url berubah / force refresh)

Arsitektur

Modul	Peran
registry.py	resolusi nama → {repo, docs_url, trust}
ingest.py	fetch → trafilatura → chunk 256 token (overlap 50)
store.py	SQLite: libs + chunks_fts (BM25) + chunks_vec (sqlite-vec)
server.py	FastMCP stdio server + --warmup CLI

Data: ~/.local/share/memo/docs.db (buat sekali, dipakai semua library).

Lisensi

MIT