autodocs-engine

Deterministic codebase intelligence for AI coding tools.

What It Does

Analyzes TypeScript/JavaScript codebases and extracts actionable intelligence — commands, conventions, file coupling patterns, contribution recipes, architecture — that AI coding tools need to write code that fits your project.

Two ways to use it:

1. MCP Server — Live codebase queries via Model Context Protocol. AI tools ask for exactly the context they need, on demand.
2. AGENTS.md Generator — Static context file for universal compatibility. Works with every tool that reads AGENTS.md, CLAUDE.md, or .cursorrules.

Quick Start

# Generate a focused AGENTS.md (no API key needed)
npx autodocs-engine init

# Or start the MCP server for live queries
npx autodocs-engine serve

MCP Server Setup (Claude Code)

claude mcp add autodocs -- npx autodocs-engine serve

This gives Claude Code 13 codebase intelligence tools:

Tool	What It Returns
get_commands	Build, test, lint commands with exact flags
get_architecture	Directory structure, entry points, package type
get_conventions	DO/DON'T rules with confidence levels (filterable by category)
get_workflow_rules	File coupling and co-change patterns (filterable by file)
get_contribution_guide	How to add new code, with inline example snippets
get_exports	Public API with usage examples and parameter shapes
analyze_impact	Blast radius + importers + co-change partners for a file
list_packages	Monorepo package inventory
plan_change	Full change plan: dependents, co-changes, registrations, tests, checklist
get_test_info	Test file path + exact per-file run command for any source file
auto_register	Exact code insertions for registration files + barrel updates
review_changes	Pattern compliance check: suffix, imports, registration, barrel, tests
diagnose	Root cause analysis: traces test failures to suspect files via import graph, git co-change, and call graph

Why Minimal Mode?

Research shows comprehensive context files can actually hurt AI performance:

• LLM-generated AGENTS.md: -0.5% to -2% accuracy (arxiv 2602.11988)
• Developer-written focused AGENTS.md: +4% accuracy, -29% runtime (arxiv 2601.20404)

The difference: developer-written files are short, focused, and only include what the AI can't figure out on its own. That's exactly what --minimal generates.

Minimal mode (~200-450 tokens) includes only:

• Exact commands with flags (proven to help, never hurts)
• Workflow rules from git co-change analysis (patterns the AI can't discover from code)
• High-confidence conventions (only if ≥95% consistent and non-obvious)
• Non-obvious directories (with "non-exhaustive" qualifier to prevent anchoring)

Full mode (~1,500-2,500 tokens) adds public API, dependency graphs, all conventions, and architecture details. Use --full when you need comprehensive documentation.

How It Works

Unlike tools that dump code into an LLM or pack everything into one file, autodocs-engine uses deterministic static analysis:

1. Parse — AST analysis via TypeScript Compiler API
2. Detect — 9 convention detectors (naming, hooks, testing, import ordering, frameworks, etc.)
3. Extract — Commands from package.json, turbo.json, biome.json, and 10+ config formats
4. Graph — Call graph, import chains, and git co-change analysis (Jaccard similarity)
5. Score — Inferability scoring decides what's worth including vs. what the AI already knows
6. Generate — 14/16 sections are deterministic (no LLM). Only architecture summary and domain terms use optional micro-LLM calls.

No API key needed for minimal mode or JSON output. The analysis is pure computation.

Output Formats

npx autodocs-engine init                         # Focused AGENTS.md (~300 tokens, no API key)
npx autodocs-engine init --full                  # Comprehensive AGENTS.md (needs API key)
npx autodocs-engine analyze . --format json      # Raw analysis JSON
npx autodocs-engine analyze . --format claude.md # CLAUDE.md format
npx autodocs-engine analyze . --format cursorrules # .cursorrules format

Monorepo Support

Auto-detects workspace packages from pnpm-workspace.yaml, workspaces field in package.json, turbo.json, or nx.json:

npx autodocs-engine init    # Works for monorepos too

For explicit control:

npx autodocs-engine analyze packages/app packages/hooks packages/ui \
  --format agents.md --hierarchical --root .

Staleness Detection

Check if your AGENTS.md needs updating (for CI):

npx autodocs-engine check

Returns exit code 1 if conventions have drifted. Useful in CI pipelines to keep context files honest.

Tested On

Repo	Files	Time	Token Count (minimal)
autodocs-engine	115	450ms	~443
vitest	1,200+	1.2s	~307
nitro	469	220ms	~121
sanity	3,746	1.6s	—
medusa	720	316ms	—

569 tests. Zero type errors. 13 MCP tools. Zero technology hallucinations across all tested repos.

Library API

import { analyze, generateMinimalAgentsMd } from 'autodocs-engine';

// Analyze (pure computation, no API key)
const analysis = await analyze({ packages: ['./packages/my-pkg'] });

// Generate minimal AGENTS.md (no API key)
const minimal = generateMinimalAgentsMd(analysis);

// Or use the full deterministic format (optional API key for 2 LLM sections)
import { formatDeterministic } from 'autodocs-engine';
const full = await formatDeterministic(analysis, config);

Configuration

Optional autodocs.config.json:

{
  "exclude": ["**/vendor/**", "**/generated/**"],
  "conventions": { "disable": ["telemetry-patterns"] }
}

Most options are auto-detected. Zero config is the default.

CLI Reference

autodocs-engine init [--full]              Generate AGENTS.md (minimal by default, no API key)
autodocs-engine serve [path]               Start MCP server
autodocs-engine check                      Check if AGENTS.md needs regeneration
autodocs-engine analyze [paths...] [options]

Options:
  --full             Comprehensive output (requires ANTHROPIC_API_KEY)
  --minimal          Focused output (<500 tokens, no API key needed — default for init)
  --telemetry        Enable session telemetry (writes to ~/.autodocs/telemetry/)
  --format, -f       json | agents.md | claude.md | cursorrules
  --output, -o       Output directory (default: .)
  --root             Monorepo root directory
  --hierarchical     Root + per-package output
  --merge            Preserve human-written sections when regenerating
  --verbose, -v      Timing details
  --dry-run          Print to stdout (no file writes)

Contributing

git clone https://github.com/msiric/autodocs-engine.git
cd autodocs-engine
npm install
npm test          # 569 tests
npm run typecheck # Zero errors
npm run build

License

MIT