Every AI agent that touches the web today is duct-taped together from six different vendors — a browser driver, a perception layer, a stealth proxy, a captcha solver, an auth manager, a benchmark harness — and burns ~2 million tokens to do what a human does in 30 seconds.

We think that's nuts.

genATP is one install that replaces the whole stack — and adds the two things nothing else has: deterministic, self-healing Action Cassettes that replay with zero LLM calls, and a declarative ATP Spec that doubles as test, automation, and production agent tool.

The web speaks human. Agents speak structured data. We're the translator everybody else forgot to build.

1. A web-access tool that needs six vendors is broken. It should be one install.
2. Tokens are a real cost. Every snapshot, every prompt, every retry — measured, compressed, justified.
3. Workflows shouldn't pay the LLM tax twice. Record once, replay forever.
4. Determinism is a feature. Agents that drift silently are worse than agents that fail loudly.
5. The same spec should run as a test, an automation, and an agent tool. Three formats for one workflow is two too many.
6. Source-available beats closed-source for trust, and trust beats marketing.

# Pick your flavor
pip install genatp        # Python
npm install genatp        # TypeScript / Node
brew install genatp       # CLI on macOS

genatp init               # one-time setup

Then point your agent — Claude Desktop, Cursor, LangChain, your own — at the MCP server, SDK, or HTTP endpoint. Full quickstart in the project README.

Full phased plan: docs/DEVELOPMENT_PLAN.md.

• ⭐ Star genatp/genatp — it genuinely helps
• 📖 Read the PRD before opening an issue — most "feature ideas" are already answered
• 🐛 File an issue with a failing repro for bugs, with a use case for proposals
• 🤝 Open a PR — see CONTRIBUTING.md
• 🔐 Report a security issue privately — see SECURITY.md