Open-source code intelligence infrastructure. Builds a programmable Intelligence Model from any codebase — queryable by AI agents, your team, or your own tooling.
Artifactor combines tree-sitter static analysis with LLM inference to produce a structured Intelligence Model: a knowledge graph of entities and relationships, a reasoning graph of purposes and business rules, and 13 documentation sections — each backed by verifiable citations to your source code. It runs on your infrastructure, stores nothing in the cloud, and works with any LLM backend.
Point Artifactor at a repository. It builds three layers:
A structured knowledge graph and reasoning graph built from dual-path analysis (AST + LLM cross-validation). Entities, relationships, call graphs, dependency graphs, business rules, and risk patterns — all with dynamic confidence scores (0-100%). Every finding cites the exact file, function, and line. Low-confidence claims are flagged, not hidden.
An MCP server (10 tools, 5 prompts) that feeds your Intelligence Model to AI coding agents — Claude, Cursor, Copilot, or any MCP-compatible tool. A REST API (30 endpoints) for programmatic access. Intent-based chat routing dispatches to specialist agents (lookup, code exploration, search) with only the relevant tools. Five built-in playbooks for guided workflows: code review, bug diagnosis, test generation, migration planning, and repo onboarding.
13 citation-grounded sections covering executive overview, system architecture, data models, API specifications, features, user stories, security analysis, and more. Exported as Markdown, HTML, PDF, or JSON.
Two independent analysis paths run against your code: static analysis (tree-sitter AST parsing) extracts structure deterministically, while LLM inference extracts meaning probabilistically. Results are cross-validated using token-based matching — entities confirmed by both paths get high confidence, single-source findings get lower scores. Citations are verified against the source tree before output.
For the full technical deep-dive, see ARTIFACTOR.md.
- Python 3.12+
- An LLM API key (OpenAI by default, or any litellm-supported provider)
uv sync
export OPENAI_API_KEY=your-key-here
artifactor analyze /path/to/your/repoExport in other formats:
artifactor analyze /path/to/repo --format html -o report.html
artifactor analyze /path/to/repo --format pdf -o report.pdf
artifactor analyze /path/to/repo --format json -o report.jsondocker compose up # Build + run (backend:8000, frontend:3000)Open http://localhost:3000 for the UI. The frontend proxies /api/* to the backend. Data persists in ./data via mounted volumes.
For standalone Docker, environment variables, and details, see ARTIFACTOR.md → Docker.
./dev.sh # Start backend + frontend (Ctrl-C to stop both)Or manually:
uv run uvicorn artifactor.main:app --reload # Backend (port 8000)
cd frontend && pnpm install && pnpm dev # Frontend (port 3000)See CONTRIBUTING.md for tests, linting, and PR guidelines.
For comprehensive technical documentation, see ARTIFACTOR.md:
- REST API (30 endpoints) and MCP server (10 tools, 5 prompts)
- Chat & RAG pipeline with hybrid vector + keyword search
- Architecture patterns and design axioms
- All 13 section generators and export formats
- Configuration reference (environment variables)
- Project structure and development workflow
- Infrastructure, not a destination — MCP server, REST API, and playbooks. Not a web viewer you visit — infrastructure you build on.
- Your code, your Intelligence Model — runs on your machine, no telemetry, no cloud storage, read-only analysis. Your code never leaves your infrastructure. See Security Policy.
- Pluggable LLM — any litellm-supported provider works: OpenAI, Anthropic, Google, Azure, local Ollama. No vendor lock-in.
- Verified citations — every claim cites the exact file, function, and line. Dual-path cross-validation (AST + LLM) catches what either path alone would miss.
- Honest confidence — every finding carries a confidence score. Low-confidence claims are flagged, not hidden. Artifactor tells you what it doesn't know.
- Engineering rigor — 696+ tests across a 4-tier test pyramid, strict type checking, protocol-based architecture, frozen value objects.
Built by Patrick Pena, Agenisea™ 🪼