sourcebook

AI can read your code. It still doesn't know how your project works.

sourcebook captures the project knowledge your team carries in its head — conventions, patterns, traps, and where things actually go — and turns it into context your coding agent can use.

npx sourcebook init

Tools like Repomix give AI your entire codebase. sourcebook gives it your project knowledge.

Why

AI coding agents spend most of their context window orienting — reading files to build a mental model before doing real work. Most context files (CLAUDE.md, .cursorrules) are generic and go stale fast.

Research shows auto-generated context that restates obvious information actually makes agents worse by 2-3%. The only context that helps is non-discoverable information — the project knowledge agents can't figure out by reading code alone.

sourcebook extracts only what agents keep missing: the conventions, hidden dependencies, fragile areas, and dominant patterns that live in your team's heads — not in the code.

What It Finds

• Import graph + PageRank — ranks files by structural importance, identifies hub files with the widest blast radius

• Git history forensics — reverted commits ("don't do this" signals), co-change coupling (invisible dependencies), rapid re-edits (code that was hard to get right), anti-patterns from abandoned approaches

• Convention detection — naming patterns, export style, import organization, barrel exports, path aliases, type hint usage, error handling style

• Framework detection — Next.js, Expo, Supabase, Tailwind, Express, TypeScript, Django, FastAPI, Flask, Go (Gin, Echo, Fiber)

• Context-rot-aware formatting — critical constraints at the top, reference info in the middle, action prompts at the bottom (optimized for LLM attention patterns)

• Smart budget enforcement — when context exceeds your token budget, drops low-priority sections first (keeps critical constraints always)

Quick Start

# Generate CLAUDE.md + AGENTS.md for your project
npx sourcebook init

# Generate for a specific tool
npx sourcebook init --format claude,agents  # CLAUDE.md + AGENTS.md (default)
npx sourcebook init --format cursor         # .cursor/rules/sourcebook.mdc + .cursorrules
npx sourcebook init --format copilot        # .github/copilot-instructions.md
npx sourcebook init --format agents         # AGENTS.md only
npx sourcebook init --format all            # All of the above

# Re-analyze while preserving your manual edits
npx sourcebook update

# See what changed since last generation (exit code 1 = changes found)
npx sourcebook diff

# Limit output to a token budget (drops low-priority sections first)
npx sourcebook init --budget 1000

Commands

Options

Language Support

GitHub Action

Auto-update context files on every merge:

# .github/workflows/sourcebook.yml
name: Update context files
on:
  push:
    branches: [main]

jobs:
  sourcebook:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: maroondlabs/sourcebook@main
        with:
          format: all

Example Output

Running on cal.com (10,456 files):

sourcebook
Extracting repo truths...

✓ Scanned project structure
  10,456 files, 3 frameworks detected
✓ Extracted 11 findings

  ● Core modules: types.ts imported by 183 files — widest blast radius
  ● Circular deps: bookingScenario.ts ↔ getMockRequestData.ts
  ● Co-change: auth/provider.ts ↔ middleware/session.ts (88% correlation)
  ● Dead code: 1,907 orphan files detected
  ● Conventions: named exports preferred (26:2 ratio)
  ● Barrel exports: 40 index.ts re-export files
  ● Commit style: Conventional Commits (feat/fix/docs)

✓ Wrote CLAUDE.md
✓ Wrote AGENTS.md

How It Works

sourcebook runs five analysis passes, all deterministic and local — no LLM, no API keys, no network calls:

1. Static analysis — framework detection, build commands, project structure, environment variables

2. Import graph — builds a directed graph of all imports, runs PageRank to find the most structurally important files

3. Git forensics — mines commit history for reverts, anti-patterns, co-change coupling, churn hotspots, and abandoned approaches

4. Convention inference — samples source files to detect naming, import, export, error handling, and type annotation patterns

5. Budget enforcement — if output exceeds your token budget, intelligently drops low-priority sections (supplementary findings first, critical constraints never)

Then applies a discoverability filter: for every finding, asks "can an agent figure this out by reading the code?" If yes, drops it. Only non-discoverable information makes it to the output.

Output is formatted for context-rot resistance — critical constraints go at the top and bottom of the file (where LLMs pay the most attention), lightweight reference info goes in the middle.

MCP Server

sourcebook serve starts a local MCP (Model Context Protocol) server that exposes live codebase intelligence to any MCP-compatible AI client — Claude Desktop, Cursor, and others.

Instead of a static context file, your AI agent can query your project's architecture on demand: look up blast radius before editing, check conventions before writing code, mine git history for anti-patterns.

Installation

Add sourcebook to your MCP client config:

{
  "mcpServers": {
    "sourcebook": {
      "command": "npx",
      "args": ["-y", "sourcebook", "serve", "--dir", "/path/to/your/project"]
    }
  }
}

Claude Code — run in your terminal:

claude mcp add sourcebook -- npx -y sourcebook serve --dir /path/to/your/project

Or add manually to ~/.claude/claude_desktop_config.json.

Claude Desktop — add to ~/Library/Application Support/Claude/claude_desktop_config.json.

Cursor — add to .cursor/mcp.json in your project or ~/.cursor/mcp.json globally.

Other MCP clients — any client that supports STDIO transport works with the same config block above.

Restart your client after updating the config.

Available Tools

The server caches the scan in memory — subsequent tool calls are fast. Pass refresh: true to analyze_codebase to force a re-scan.

Roadmap

• .cursor/rules/sourcebook.mdc + legacy .cursorrules output

• .github/copilot-instructions.md output

• sourcebook update — re-analyze while preserving manual edits

• sourcebook diff — show what changed (CI-friendly exit codes)

• --budget <tokens> — smart PageRank-based prioritization

• Anti-pattern detection from reverted commits and deleted files

• Python support (Django, FastAPI, Flask, pytest)

• Go support (Gin, Echo, Fiber, module layout)

• GitHub Action for CI

• sourcebook serve — MCP server mode

• Framework knowledge packs (community-contributed)

• Tree-sitter AST parsing for deeper convention detection

• Hosted dashboard with context quality scores

Research Foundation

Built on findings from:

• ETH Zurich AGENTS.md study — auto-generated obvious context hurts agent performance

• Karpathy's autoresearch — curated context (program.md) is the #1 lever for agent effectiveness

• Aider's repo-map — PageRank on import graphs for structural importance

• Chroma's context-rot research — LLMs show 30%+ accuracy drops for middle-of-context information

License

BSL-1.1 — source-available, free to use, cannot be offered as a hosted service. Converts to MIT on 2030-03-25. See LICENSE for details.