Claude Code

June 23, 2026 · View on GitHub

Docs › Harnesses › Claude Code

Claude Code

Tier: Fully supported · Repo: a5c-ai/babysitter-claude · Harness key: claude-code

In Plain English

On Claude Code you get the full Babysitter experience: the /babysitter:* slash-commands and the babysit skill, driven by Claude Code's session and Stop hook.

Install the plugin once, restart Claude Code, and run /babysitter:call <your request>. Babysitter orchestrates the run from inside your Claude Code session - one orchestration phase per turn - and pauses for your approval at breakpoints right in the chat.

Estimated time to first run: about 5 minutes. End state: you can type /babysitter:call in Claude Code and watch a run iterate to a quality target.

On this page

Install

Install the core CLI, then add the Claude Code plugin from the marketplace:

# 1. Core CLI (host-side)
npm install -g @a5c-ai/babysitter@latest

# 2. Add the Babysitter marketplace and install the plugin
claude plugin marketplace add a5c-ai/babysitter-claude
claude plugin install --scope user babysitter@a5c.ai
claude plugin enable --scope user babysitter@a5c.ai

Then restart Claude Code so the plugin loads.

The marketplace add uses the released default branch - do not pin a staging ref. The Claude marketplace publishes its own released 5.x version; you do not need to (and should not) specify a -staging.<sha> build.

Alternative (SDK helper): the universal installer wires up the same plugin from the core CLI:

babysitter harness:install-plugin claude-code [--workspace <path>]

Verify

After restarting Claude Code, confirm the skill is available:

/skills

You should see babysit (and the /babysitter:* commands) in the list. If not, run claude plugin list to confirm the plugin is installed and enabled, then restart Claude Code again.

Command Surface

Claude Code exposes Babysitter as first-class /babysitter:<command> slash-commands plus the babysit skill:

/babysitter:call          /babysitter:plan          /babysitter:yolo
/babysitter:forever       /babysitter:resume        /babysitter:doctor
/babysitter:retrospect    /babysitter:assimilate    /babysitter:cleanup
/babysitter:observe       /babysitter:contrib       /babysitter:help
/babysitter:blueprints    /babysitter:user-install  /babysitter:project-install
babysit                   (skill - verify via /skills)

See Slash Commands and Modes for what each mode does.

Hook / Continuation Model

Model: SessionStart + Stop (one orchestration phase per turn).

  • SessionStart prepares state and environment for the session.
  • The assistant performs one orchestration phase per turn, posts its effect, and stops.
  • The synchronous Stop hook then decides block-vs-approve-exit: it blocks (and injects the next iteration) while the run is in progress, and only allows the session to finish when the completion proof returns as <promise>...</promise>.

This Stop-hook continuation is specific to Claude Code. Other harnesses use different models (AfterAgent, daemon agent_end, ACP, session-idle, thin-skill aliases) - see Hooks and the Install Matrix. Do not assume the Claude Stop model elsewhere.

First Run

/babysitter:call build a calculator with add, subtract, multiply, divide using TDD

Babysitter creates a run, iterates (write → test → fix) toward your quality target, and pauses for approval at any breakpoints - all inside the chat. Resume an interrupted run with /babysitter:resume.

Next steps