Babysitter 🤖 Agent Open source

What it does

Babysitter is an orchestration framework that makes coding agents follow a workflow defined in code rather than improvising. You describe a process in JavaScript, or let the agent plan one, and Babysitter runs it step by step: it enforces quality gates before moving on, pauses at breakpoints for human approval, and records every decision in an event-sourced journal so interrupted runs can be resumed or reviewed. The project's own wording promises obedience and "hallucination-free" orchestration; in practice it gives you deterministic control flow around an agent that can still make mistakes inside each step.

What is inside

• In-session commands such as /babysitter:call (interactive with approvals), /babysitter:plan, /babysitter:yolo (no breakpoints), /babysitter:forever for recurring tasks, plus doctor, observe and resume
• The babysitter CLI and SDK, which manage runs, session state and harness plugins
• An optional genty runtime CLI for running processes from the shell or CI, including an internal harness that needs no external coding agent
• An adapters CLI that installs and drives supported harnesses from one place
• Process blueprints, a knowledge catalog package, a monitoring dashboard and an MCP server

Works with

Claude Code has the most complete support, followed by Codex (beta). Cursor, Gemini CLI, GitHub Copilot, OpenCode and several pi-based harnesses are listed as experimental.

How to install

npm install -g @a5c-ai/babysitter

For Claude Code, add the plugin:

claude plugin marketplace add a5c-ai/babysitter-claude
claude plugin install --scope user [email protected]

Maintenance and safety

The project is MIT-licensed, published on npm and very actively developed, with CI and extensive documentation. It is a sizeable monorepo with several packages, so expect a learning curve and frequent changes. Autonomous and continuous modes run without approval gates; use interactive mode for anything touching production. Run journals are stored in your workspace under .a5c.

Who should use it

Developers and teams who need repeatable, auditable multi-step agent runs, such as TDD feature delivery, migrations or CI automation, and are willing to define processes up front. For casual single-prompt work it is more machinery than needed.