berriai/moyai is a Python-based, self-hosted coding agent. It works with more than 100 providers, including Claude Code, Codex and OpenCode, and you can instruct it in conversation and steer it while it runs. You can also inspect the resulting files and the task history. This repository is a project introduction, not a new release.
The core is a setup that puts durable execution — which can pick a task back up after it is interrupted — and an isolated agent computer into a single workspace. Durable execution uses Temporal, and the agent computer runs on Modal by default. Model integration uses LiteLLM, while the agent layer uses Pydantic AI and a harness. The repository also includes real task items and grading criteria, so the evaluation structure can be examined. That said, first-party benchmark results such as performance numbers or pass rates could not be confirmed.
Self-hosting means you operate the application yourself, not that every task runs on your local machine. The agent computer runs on Modal by default, and a team deployment requires a database and a private object store, among other things. The 5-out-of-5 pass result in the documentation predates the desktop adapter, so it does not demonstrate the current adapter's performance. The setup and the evaluation path can therefore be reviewed, but real-world results are still hard to judge.