Drizzle ORM文档检索MCP服务

drizzle-docs-mcp

Your AI writes Drizzle code from memory. Memory is out of date, so you get code that looks right and doesn't run — a where clause that was renamed two versions ago, an import that moved, a Postgres trick that only works on SQLite.

This gives it the real thing instead. 446 pages across 6 SQL dialects, read straight from orm.drizzle.team and served as clean Markdown — so the answer comes from today's docs, not from a training set.

npx -y drizzle-docs-mcp

That's the whole install. No API key, no account, nothing to configure.

What you get

Three tools your AI can call:

• depth: "index" (default) fuzzy-searches titles, slugs and sections straight from llms.txt — instant, and it tolerates typos (migraton finds Migrations).
• depth: "full" downloads llms-full.txt once (a few MB) and searches the actual page text, returning a snippet around each match. Use it when you need the answer, not a link.
• Slugs look like docs/pg/select. Some catalogue links point at pages the site serves without the dialect segment (/docs/pg/overview → /docs/overview) — fetch_page retries the unprefixed URL and repairs the catalogue, so you never have to care.

Ask it things like:

• "How do I set up a Postgres schema in Drizzle?"
• "Which migration pages exist for SQLite?"
• "Show me the batch API examples" → search_docs { query: "batch api", depth: "full" }

The catalogue comes from Drizzle's own llms.txt and llms-full.txt, so it tracks the real docs structure instead of scraping a sidebar and guessing.

Connect

1. npm (stdio) — nothing to install

npx -y drizzle-docs-mcp        # Node
bunx drizzle-docs-mcp          # Bun

Or add it to your editor's MCP config:

{
  "mcpServers": {
    "drizzle-docs": {
      "command": "npx",
      "args": ["-y", "drizzle-docs-mcp"]
    }
  }
}

{
  "drizzle-docs": {
    "command": "npx",
    "args": ["-y", "drizzle-docs-mcp"]
  }
}

{
  "mcpServers": {
    "drizzle-docs": { "command": "npx", "args": ["-y", "drizzle-docs-mcp"] }
  }
}

{
  "servers": {
    "drizzle-docs": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "drizzle-docs-mcp"]
    }
  }
}

claude mcp add drizzle-docs -- npx -y drizzle-docs-mcp

{
  "context_servers": {
    "drizzle-docs": {
      "command": { "path": "npx", "args": ["-y", "drizzle-docs-mcp"] }
    }
  }
}

2. https (Streamable HTTP) — host it yourself

The same server can listen on a port. Set MCP_TRANSPORT=http (or pass --http) and the MCP endpoint is served at /mcp:

docker build -t drizzle-docs-mcp .
docker run --rm -p 3000:3000 drizzle-docs-mcp
curl http://localhost:3000/        # {"name":"drizzle-docs-mcp","version":"…",…}

Without Docker:

bun install && bun run build
node dist/index.js --http          # PORT, HOST, MCP_PATH come from the env

Then point a client at it:

{
  "mcpServers": {
    "drizzle-docs": { "type": "http", "url": "https://<your-host>/mcp" }
  }
}

Runs on Node 22+ or Bun, so any container platform, VM or PaaS works — tmcp speaks plain Web Request/Response, which is why the same code serves stdio locally and HTTP remotely.

[!NOTE] v2 ran on Mastra Cloud (drizzle.mastra.cloud). That endpoint is retired — use npx for local use, or self-host for a shared URL. See CHANGELOG.md.

Skills

The repo ships an agent skill that teaches your AI how to use these tools (when to use which one, and the gotchas above). Install it:

# From the skills registry (same pattern as DocShark)
npx skills add Michael-Obele/drizzle-docs --skill drizzle-docs

Or by hand into the standard skill folder:

mkdir -p ~/.agents/skills/drizzle-docs
curl -fsSL https://raw.githubusercontent.com/Michael-Obele/drizzle-docs/master/skills/drizzle-docs/SKILL.md \
  -o ~/.agents/skills/drizzle-docs/SKILL.md

• GitHub Copilot and OpenCode read ~/.agents/skills/ automatically.
• Claude Code: ln -s ../../.agents/skills/drizzle-docs ~/.claude/skills/drizzle-docs
• In this repo the skill lives in skills/drizzle-docs/ and is symlinked into .agents/, .agent/ and .windsurf/, so every supported editor picks it up from a clone.

Also: DocShark

This server answers one site — Drizzle's. If you work across many documentation sites, use DocShark (our own docs MCP, also built on tmcp): it crawls any docs site, stores it in SQLite with FTS5/BM25 search, and lets your AI query the latest pages.

bun add -g docshark

docshark add https://orm.drizzle.team/ --depth 2   # index a site
docshark search "relational queries"               # CLI search
docshark list                                      # what is indexed
docshark stale                                     # refresh anything >14 days old

Add it to your MCP config:

{
  "mcpServers": {
    "docshark": {
      "command": "bunx",
      "args": ["-y", "docshark", "start", "--stdio"]
    }
  }
}

Its agent skills are one command away too:

npx skills add Michael-Obele/docshark --skill docshark
npx skills add Michael-Obele/docshark --skill using-docshark

Use them together: drizzle-docs-mcp for instant, always-fresh Drizzle answers, DocShark for everything else you index. See DocShark on GitHub →.

Configuration

All optional — defaults work out of the box.

Local development

bun install          # dependencies
bun run dev          # stdio, watch mode
bun run dev:http     # HTTP on :3000/mcp
bun run check        # tsc --noEmit
bun test             # parser + search tests (offline fixtures)
bun run build        # emit dist/ (this is what npm ships)

Publishing a release

Bump version in package.json, then push a matching tag. The workflow typechecks, tests, builds, publishes to npm and cuts a GitHub release:

bun version patch        # 3.0.1 -> 3.0.2
git commit -am "chore: release"
git tag vX.Y.Z && git push origin vX.Y.Z

It refuses to publish if the tag does not match package.json, and prepublishOnly re-runs check + test + build first.

Architecture

• MCP_ARCHITECTURE.md — how the server, the data layer and the transports fit together.
• src/lib/docs.ts — catalogue (llms.txt), page fetching + Markdown conversion, full corpus (llms-full.txt), caching.
• src/lib/search.ts — Fuse.js indexes for the fast and the deep search.
• src/tools/ — the three tools.
• src/server.ts + src/index.ts — server assembly and transport selection.

Contributing

Read the Contributing Guidelines and Code of Conduct first. Issues and security reports go to Security.

License

MIT