@tesseron/docs-mcp

Tesseron documentation as an MCP server. Search and read the full protocol + SDK docs from any MCP-compatible AI client: Claude Code, Claude Desktop, Cursor, Windsurf, Zed, and anything else that speaks the Model Context Protocol.

No network, no vector store, no API keys. The docs snapshot is bundled in the npm package at publish time. Search runs locally with BM25 over in-memory text.

Why

If you are writing a Tesseron app, your AI assistant should know Tesseron. This server exposes the canonical Tesseron docs as three MCP tools (list_docs, search_docs, read_doc) plus tesseron-docs://<slug> resources so the agent can retrieve exact spec text before answering instead of guessing from training data.

Install

Nothing to install. Use npx.

Configure your AI client

Claude Code / Claude Desktop

Add to ~/.claude.json or ~/.config/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "tesseron-docs": {
      "command": "npx",
      "args": ["-y", "@tesseron/docs-mcp"]
    }
  }
}

Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "tesseron-docs": {
      "command": "npx",
      "args": ["-y", "@tesseron/docs-mcp"]
    }
  }
}

Windsurf / Zed / any other MCP client

Point the client at npx -y @tesseron/docs-mcp. stdio transport, Node ≥ 20.

Tools

| Tool | Input | Returns | |---|---|---| | list_docs | {} | { count, docs: Array<{ slug, title, section, description, related }> } | | search_docs | { query: string, limit?: number (1-20, default 8) } | { query, count, hits: Array<{ slug, title, description, section, score, snippet }> } | | read_doc | { slug: string } | { slug, title, description, section, related, body } (body is full markdown) |

Slug format: <section>/<basename> without extension. Examples: protocol/handshake, sdk/typescript/action-builder, examples/react-todo.

Resources

Each docs page is also exposed as an MCP resource at tesseron-docs://<slug>. Clients that prefer the resource surface over tools (Claude Desktop is one) get full listing via list_resources and full markdown bodies via read_resource.

Search behaviour

• BM25 index via minisearch.
• Fields: title (weight 3), description (weight 2), bodyText (weight 1).
• Fuzzy matching (0.15) and prefix matching are both on.
• Snippets are ~240 chars centred on the best match, with ellipses added when truncated.

Dev / local use

Point the server at a local snapshot or set the env var:

# Build fresh snapshot from the monorepo docs tree.
pnpm --filter @tesseron/docs-mcp build:snapshot

# Run from source with a custom snapshot file.
pnpm --filter @tesseron/docs-mcp start -- --snapshot ./dist/docs-index.json

# Or via env var.
TESSERON_DOCS_SNAPSHOT=/abs/path/to/docs-index.json npx @tesseron/docs-mcp

The snapshot is rebuilt on every pnpm build (via pnpm build:snapshot && tsup) and on prepublishOnly.

Versioning

@tesseron/docs-mcp versions track the docs snapshot, not the SDK. A new publish means the bundled docs were updated. It is unscoped from the @tesseron/{core,web,server,react,mcp} fixed-version bundle so SDK releases do not churn the docs package.

License

BUSL-1.1. See LICENSE.