doctree-mcp

Agentic document retrieval over markdown, CSV, and JSONL. BM25 + tree navigation via MCP — no vector DB, no embeddings, no LLM calls at index time.

The pitch: MCP provides the structural primitives (a navigable tree, BM25, glossary, row lookup). The bundled skills provide the procedural knowledge (how to walk that tree). Together the agent behaves like a trained research librarian — not a one-shot searcher. See The Skill + MCP Pattern.

Quick Start

Have docs already? Point a client at them:

# In your AI tool's MCP config — see docs/CLIENTS.md for per-tool snippets
{ "mcpServers": { "doctree": {
    "command": "bunx", "args": ["doctree-mcp"],
    "env": { "DOCS_ROOT": "./docs", "WIKI_WRITE": "1" }
} } }

Restart the tool → ask "search the docs for X" or invoke the doc-read prompt.

Starting fresh? Scaffold a Karpathy-style LLM wiki:

bunx doctree-mcp init          # configure current tool
bunx doctree-mcp init --all    # configure every supported client
bunx doctree-mcp init --dry-run

Creates docs/wiki/ (LLM-maintained) + docs/raw-sources/ (your inputs), writes the MCP config, installs a post-write lint hook, appends wiki conventions to CLAUDE.md / AGENTS.md / .cursor/rules/.

Related MCP server: search-docs

Operation Modes

Full decision tree: Operation Modes.

How It Works — Retrieve · Curate · Add

Agent: "How does token refresh work?"

→ search_documents("token refresh")
  #1  auth/middleware.md § Token Refresh Flow       score: 12.4
  #2  auth/oauth.md       § Refresh Token Lifecycle  score: 8.7

→ get_tree("docs:auth:middleware")
  [n1] # Auth Middleware
    [n4] ## Token Refresh Flow
      [n5] ### Automatic Refresh

→ navigate_tree("docs:auth:middleware", "n4")   ← n4 + descendants

Core read tools (always on):

Wiki write tools (opt-in with WIKI_WRITE=1):

Safety: path containment · frontmatter validation · duplicate detection · dry-run · overwrite protection.

Deprecated aliases (list_documents, find_files, find_symbol) are superseded by search_documents — still functional, no longer recommended.

The Skill + MCP Pattern

Most retrieval tools hand the agent a search box and hope for the best. doctree-mcp hands it a tree, and the bundled skills teach it how to walk one.

• MCP = structural primitives. search_documents, get_tree, navigate_tree, get_node_content, lookup_row return tree positions the agent reasons over — not finished answers.

• Skills = procedural knowledge. /doc-read, /doc-write, /doc-lint encode breadcrumb drill-down: search → outline → navigate → retrieve. The agent learns the policy, not just the API.

That pairing doesn't exist cleanly elsewhere:

Why iterative retrieval wins:

• Context rot. Stuffing a 1M-token window with chunks degrades output. Breadcrumb navigation keeps working memory small.

• Auditability. search_documents → get_tree → navigate_tree → get_node_content is a replayable trail. A cosine score is not. Regulated domains can ship the former.

• Progressive disclosure. Fewer navigable primitives beat tool sprawl (cf. Cloudflare Code Mode).

Multi-instance = client-side federation. Register several doctree servers under different names; the /doc-read skill encodes the routing policy. Add or remove instances without touching the skill. See Client setup → Multi-instance routing.

The LLM Wiki Pattern

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│  Raw Sources    │     │  The Wiki        │     │  The Schema     │
│  (immutable)    │ ──→ │  (LLM-maintained)│ ←── │  (you define)   │
│  notes · logs   │     │  runbooks · refs │     │  CLAUDE.md rules │
└─────────────────┘     └─────────────────┘     └─────────────────┘

Inspired by Karpathy's LLM Wiki. Full walkthrough: docs/LLM-WIKI-GUIDE.md.

Configuration (summary)

---
title: "Descriptive Title"
description: "One-line summary — boosts ranking"
tags: [relevant, terms]
type: runbook          # runbook | guide | reference | tutorial | architecture | adr
category: auth
---

All non-reserved frontmatter fields become filter facets:

search_documents("auth", filters: { type: "runbook", tags: ["production"] })

Common env vars:

Full reference: docs/CONFIGURATION.md.

Glossary — place glossary.json in docs root for bidirectional query expansion:

{ "CLI": ["command line interface"], "K8s": ["kubernetes"] }

Acronym definitions like "TLS (Transport Layer Security)" are also auto-extracted.

Structured data — CSV/JSONL files become documents where each row is a tree node. Column roles (id, title, description, facets, URL) are auto-detected from headers. See docs/STRUCTURED-DATA.md.

Running from Source

git clone https://github.com/joesaby/doctree-mcp.git
cd doctree-mcp && bun install

DOCS_ROOT=./docs bun run serve          # stdio
DOCS_ROOT=./docs bun run serve:http     # HTTP (port 3100)
DOCS_ROOT=./docs bun run index          # CLI: inspect indexed output
bun test

Performance

Docs

Setup & operation

• Operation Modes — stdio · HTTP · CLI

• Client Setup — Claude Code · Cursor · Windsurf · Codex · OpenCode · Claude Desktop

• Deployment — Railway · Fly.io · Render · Cloudflare Containers · Docker

• Configuration — env vars, frontmatter, ranking tuning

Patterns & concepts

• LLM Wiki Guide — agent-maintained knowledge base walkthrough

• Structured Data — CSV / JSONL indexing

• Architecture & Design — BM25 internals, tree navigation

• Competitive Analysis — PageIndex, QMD, GitMCP, Context7, managed RAG

Source

• Prompts — MCP prompt templates

• Skills: /doc-read · /doc-write · /doc-lint

Standing on Shoulders

• PageIndex — hierarchical tree navigation

• Pagefind by CloudCannon — BM25 scoring, positional index, facets

• Bun.markdown by Oven — native CommonMark parser

• Karpathy's LLM Wiki — the LLM-maintained wiki pattern

License

MIT