What it is
Babysitter is a monorepo of CLIs, SDKs and per-harness plugins that governs AI coding agents with processes written as JavaScript code. The orchestrator can only do what the process code permits, and every step ends with a forced stop. Quality gates block progression, breakpoints require human approval, and every decision is recorded in an event-sourced journal that supports replay and resume. As of v6 it is harness-agnostic through its Adapters runtime and supports 12 AI coding harnesses.
Who it's for
- Developers who want AI coding agents to follow deterministic, code-defined workflows
- Teams that need human approval breakpoints and an audit trail of agent decisions
- CI/CD and scripting users who want headless orchestration without an external AI coding agent
Requirements
Requirements
- Node.js 20.0.0+ (22.x LTS recommended)
- The host-side
adaptersCLI requires Node 22.13.0+ - A supported AI coding harness from the 12 in the install matrix (e.g. Claude Code)
- Git (optional, for cloning)
Setup
Install the main CLI
Recommended end-user install of the core
babysitterCLI.bashnpm install -g @a5c-ai/babysitterInstall the host-side adapters CLI (optional)
Run any supported harness directly from your shell.
bashnpm install -g @a5c-ai/adapters-cli adapters doctor adapters run claude "explain this codebase"Install the Claude Code plugin
Native marketplace install; then restart Claude Code and type
/skillsto verify "babysit" appears.bashclaude plugin marketplace add a5c-ai/babysitter-claude claude plugin install --scope user babysitter@a5c.aiFirst-run configuration
Create your profile once, set up the project, then verify with diagnostics.
bash/babysitter:user-install /babysitter:project-install /babysitter:doctor
Examples
Quick start with Claude
bashclaude "/babysitter:call implement user authentication with TDD"What it does: Starts an interactive orchestration run that executes tasks step-by-step, handles quality checks and approvals, and continues until completion.
Run a process with the internal harness
bashnpm install -g @a5c-ai/genty-platform
genty call --harness internal --process .a5c/processes/my-process.js#process --workspace .What it does: Runs a process definition headlessly with the built-in engine, with no external AI coding agent required.
CI pipeline invocation
bashgenty call \
--harness internal \
--process .a5c/processes/lint-and-test.js#process \
--workspace . \
--no-interactive \
--jsonWhat it does: Non-interactive JSON-output run suitable for CI/CD pipelines and scripts.
Autonomous and resume runs
bashgenty yolo --harness claude-code --prompt "add pagination to the API" --workspace .
genty resume --run-id <runId> --harness claude-code --workspace .What it does: Runs fully autonomously with no breakpoints, then shows how to resume an interrupted run by its run ID.
Pros & cons
Pros
- Pro:Processes are real JavaScript code, so the orchestrator can only do what the code permits, with mandatory stops and gates that block progression
- Pro:Event-sourced immutable journal supports deterministic replay, resume and auditing
- Pro:Harness-agnostic across 12 supported AI coding harnesses, with an internal harness for headless CI/CD use
- Pro:Human-in-the-loop breakpoints plus multiple modes (interactive, autonomous, planning, continuous)
Cons
- Con:Install is split across several packages (babysitter, adapters-cli, genty-platform, per-harness plugins), which the README warns not to conflate
- Con:Most non-Claude harness plugins are marked Experimental or Beta
- Con:Node version floors differ: the adapters CLI needs 22.13.0+ while the core needs 20+
Images
