owainlewis/machinist

What it solves

Machinist provides a controlled, repeatable, and scalable environment for AI coding workflows. It prevents the risks associated with giving AI agents arbitrary shell access by restricting them to named commands and specific repositories, while ensuring that humans remain the final authority on what code is merged via pull requests.

How it works

Machinist acts as a software factory that maps specific command names to fixed executables (such as Codex, Claude Code, or custom shell scripts). When a command is run, Machinist renders a prompt, sends it to the executable via standard input, and manages the execution process—including streaming output, tracking token usage, and enforcing timeouts. All credentials and repository access remain local to the worker machine.

Who it’s for

It is designed for developers and teams using agentic coding tools who want a standardized way to orchestrate AI-driven development without sacrificing security or local control of credentials.

Highlights

• Controlled Entrypoints: Restricts agents to approved commands rather than arbitrary shell text.
• Local Authority: Keeps repositories, credentials, and model aliases on the worker machine.
• Bring Your Own Harness: Compatible with any executable that accepts prompts via standard input.
• Full Observability: Streams output and retains durable events, artifacts, and token usage data.
• Human-in-the-loop: Ensures the process ends with a pull request for human review.