Maat Agent

Maat Agent is a coding assistant for developers who want testable evidence before accepting AI edits. It runs in a terminal or an Electron desktop app on macOS, Windows and Linux, and checks the actual project files against acceptance criteria set before work begins. Failed checks send their output back to the model for another attempt; if the work still fails, the agent reports failure.

Built on the molt engine, it's open source under Apache 2.0. You can connect local models through Ollama, llama.cpp or vLLM, or use cloud providers through an OpenAI-compatible API or Anthropic's native API. The agent and its verification records run locally; choosing a cloud provider sends model requests to that service using your own API key.

Its checks go beyond a passing test suite. They can detect deleted assertions, changes outside recorded tool writes, and added code that tests never execute. Mutation checks deliberately break changed code to see if the tests catch it. These checks establish specific properties of the work, rather than guaranteeing that every result is correct.

Each completion attempt leaves a receipt, including rejected attempts, with check results and file-change evidence. Hash-linked journals let you detect alterations to the record. The desktop also exposes model exchanges and receipts for inspection.

Disputed results go to human review, and the record distinguishes a human ruling from machine verification. For longer tasks, missions divide work into features with fixed acceptance criteria and rerun milestone checks before accepting the combined result.