https://github.com/user-attachments/assets/8c3b0078-9396-48e8-aa43-5f40da30c20b

What is Babysitter?

Babysitter enforces obedience to agentic workforces, enabling them to manage extremely complex tasks and workflows through deterministic, hallucination-free self-orchestration. Define your workflow in code - Babysitter enforces every step, ensures quality gates pass before progression, requires human approval at breakpoints, and records every decision in an immutable journal. Your agents do exactly what the process permits, nothing more.

As of v6, Babysitter is harness-agnostic via its Adapters runtime: the same processes run across the 12 supported AI coding harnesses, so you are not locked to a single tool. See Adapters and the harness install matrix.

Prerequisites

• Node.js: Version 20.0.0+ (22.x LTS recommended). The host-side adapters CLI pins a slightly higher floor of 20.9.0+.
• A supported AI coding harness: any of the 12 harnesses covered in the install matrix (e.g. Claude Code — docs).
• Git: For cloning (optional)

Installation

Babysitter v6 has two install tracks that should not be conflated: the host-side adapters CLI for running any harness directly from your shell, and the in-session per-harness plugin for driving full orchestration runs from inside your harness. Most people want both. The package split is:

• @a5c-ai/babysitter is the recommended end-user install for the main babysitter CLI.
• @a5c-ai/adapters-cli provides the host-side adapters CLI (Node >=20.9.0) for running and managing any supported harness from your shell. See the Adapters CLI reference.
• @a5c-ai/babysitter-sdk is the public SDK/library package and the underlying implementation behind the core CLI.
• @a5c-ai/genty-platform is the optional runtime CLI for genty call, resume, start-server, tui, and other orchestration/runtime commands.
• Harness plugins such as @a5c-ai/babysitter-codex or @a5c-ai/babysitter-cursor integrate Babysitter into a specific host tool. They do not replace the core CLI packages.

For most users, install the main CLI first:

npm install -g @a5c-ai/babysitter

Prefer to drive a harness directly from your shell instead of from inside a session? Install the host-side adapters CLI and run any supported harness with one command — see the Adapters CLI reference:

npm install -g @a5c-ai/adapters-cli
adapters doctor
adapters run claude "explain this codebase"

Babysitter supports the 12 AI coding harnesses listed in the install matrix. Install the plugin for your harness of choice; the two fully-worked harnesses have dedicated pages — Claude Code and Codex. In this repository, plugins/babysitter-unified is the only maintained plugin source; harness-specific bundles are generated during build/release and are not committed:

Claude Code

Native marketplace install:

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

Restart Claude Code, then type /skills to verify "babysit" appears.

Claude Cowork can install the same plugin. For a personal install from this repo's Claude marketplace:

1. Open Claude Desktop and switch to the Cowork tab.
2. Click Customize in the left sidebar.
3. Click Browse plugins.
4. Select Personal.
5. Click +, then choose Add marketplace from GitHub.
6. Enter https://github.com/a5c-ai/babysitter.
7. Install the Babysitter plugin from that marketplace.

For Team and Enterprise org-managed installs, owners can add plugins through Organization settings > Plugins. GitHub-synced organization marketplaces require a private or internal GitHub repository, so use a private/internal fork of this repo for that flow, or upload plugin ZIPs manually.

Plugin README

Codex CLI (Beta)

Official marketplace install:

codex plugin marketplace add a5c-ai/babysitter

Plugin README

Cursor IDE and CLI (Experimental)

Via the Cursor marketplace or the SDK helper:

babysitter harness:install-plugin cursor

Plugin README

Gemini CLI (Experimental)

babysitter harness:install-plugin gemini-cli

Plugin README

GitHub Copilot (Experimental)

Via the GitHub Copilot CLI marketplace, or:

babysitter harness:install-plugin github-copilot

Plugin README

Pi (Experimental)

babysitter harness:install-plugin pi

Plugin README

Hermes (Experimental)

babysitter harness:install-plugin hermes

Plugin README

Oh-My-Pi (Experimental)

babysitter harness:install-plugin oh-my-pi

Plugin README

OpenCode (Experimental)

babysitter harness:install-plugin opencode

Plugin README

OpenClaw (Experimental)

babysitter harness:install-plugin openclaw

Plugin README

Internal Harness (No AI Coding Agent Required)

Babysitter ships with a built-in internal harness that runs processes programmatically without any external AI coding agent. This is useful for CI/CD pipelines, scripts, automated testing, and headless orchestration:

npm install -g @a5c-ai/genty-platform

# Run a process definition using the internal harness
genty call --harness internal --process .a5c/processes/my-process.js#process --workspace .

# Or run a free-form prompt
genty call --harness internal --prompt "run lint and tests" --workspace .

The internal harness uses the SDK's built-in Pi execution engine directly. It supports all capabilities (Programmatic, SessionBinding, StopHook, HeadlessPrompt) and requires no external AI harness CLI.

During process execution, the internal harness can delegate tasks to any discovered installed harness via the invoker. A process running under --harness internal can spawn subagent tasks that execute through Claude Code, Codex, Gemini CLI, or any other harness found on the system -- the SDK discovers available harness CLIs at runtime and routes task execution accordingly. This means you can orchestrate a multi-agent workflow from a single headless entry point, with different tasks delegated to whichever harness is best suited for them.

Runtime Package Builds

For the core runtime chain (@a5c-ai/babysitter-sdk, @a5c-ai/adapters, @a5c-ai/genty-core, @a5c-ai/genty-platform), use the shared workspace entrypoint from a fresh checkout:

npm ci
npm run build:runtime

build:runtime is the supported root entrypoint for release and CI validation. It builds the runtime graph in workspace order: SDK -> adapters SDK surface -> genty-core -> genty-platform.

Package-local validation is also supported:

npm run build --workspace=@a5c-ai/genty-core
npm run build --workspace=@a5c-ai/genty-platform

Those package-local builds now use tsc --build project references where the runtime packages are owned in this workspace, and they explicitly bootstrap the @a5c-ai/adapters SDK chain through the root runtime scripts. Fresh-checkout validation no longer assumes prebuilt upstream dist/ artifacts.

@a5c-ai/atlas provides the unified knowledge graph, ontology, and catalog data consumed by SDK, adapters, hooks, plugin tooling, and the catalog UI. The agent catalog surface is @a5c-ai/atlas/catalog. See packages/atlas/README.md.

CLI Walkthrough Verification

The published CLI walkthrough at docs/cli-examples.md is verified against the real repo surfaces, not a separate docs-only harness. From a fresh checkout, use:

npm ci
npm run build --workspace=@a5c-ai/babysitter-sdk
npm run docs:prepare
npm run docs:examples:smoke
npm run docs:qa

npm run docs:examples:verify runs that docs-focused flow end-to-end. The generated traceability map for the walkthrough lives at docs/generated/cli-examples-verification.md.

First Steps

After installation, set up your environment:

1. Configure Your Profile (One-Time)

/babysitter:user-install

This creates your personal profile with:

• Breakpoint preferences (how much oversight you want)
• Tool preferences and communication style
• Expertise areas for better process matching

2. Set Up Your Project

/babysitter:project-install

This analyzes your codebase and configures:

• Project-specific workflows
• Test frameworks and CI/CD integration
• Tech stack preferences

3. Verify Setup

/babysitter:doctor

Run diagnostics to confirm everything is working.

Quick Start

claude "/babysitter:call implement user authentication with TDD"

Or in natural language:

Use the babysitter skill to implement user authentication with TDD

Claude will create an orchestration run, execute tasks step-by-step, handle quality checks and approvals, and continue until completion.

Choose Your Mode

Utility Commands

Agent Runtime CLI

Beyond the in-session skill commands (/babysitter:call, etc.), Babysitter provides an optional agent runtime CLI package, @a5c-ai/genty-platform, for orchestration, session management, MCP serving, daemon utilities, and the TUI. The main babysitter CLI comes from @a5c-ai/babysitter and is backed by @a5c-ai/babysitter-sdk; it keeps the core run/task/session/plugin surfaces plus harness:install and harness:install-plugin.

npm install -g @a5c-ai/babysitter
npm install -g @a5c-ai/genty-platform

Running Processes via a Harness

# Run a process interactively via Claude Code (pauses at breakpoints)
genty call --harness claude-code --prompt "implement user authentication with TDD" --workspace .

# Run fully autonomous (no breakpoints)
genty yolo --harness claude-code --prompt "add pagination to the API" --workspace .

# Plan only (stops after Phase 1)
genty plan --harness claude-code --prompt "implement feature X"

# Run with the internal harness (no external AI agent needed)
genty call --harness internal --prompt "run lint and tests" --workspace .

Managing Runs

# Resume an interrupted run
genty resume --run-id <runId> --harness claude-code --workspace .

# Initialize or inspect orchestration session state
babysitter session:init --session-id demo --state-dir .a5c --run-id <runId>
babysitter session:state --session-id demo --state-dir .a5c

# Start the MCP server owned by the agent runtime CLI
genty start-server --transport stdio

# Diagnose run health
genty doctor --run-id <runId>

# Analyze past runs for insights
genty retrospect --all --harness claude-code --workspace .

# Clean up old runs
genty cleanup --keep-days 7 --harness claude-code --workspace .

Harness Discovery

# Install an agent harness CLI (preferred)
adapters install claude-code

# Install a Babysitter harness plugin globally
babysitter harness:install-plugin claude-code

# Install a Babysitter harness plugin into a workspace
babysitter harness:install-plugin codex --workspace /path/to/repo

Note: babysitter harness:install and babysitter harness:discover are deprecated. Use adapters install <agent> for agent installation and genty for runtime commands.

harness:install-plugin is the canonical scriptable install path for Babysitter plugins. For non-Claude harnesses it resolves to the published package installer shape tested in the SDK, for example npx --yes @a5c-ai/babysitter-codex install --workspace /path/to/repo.

Using --harness internal for Automation

The internal harness is particularly useful for CI/CD and scripting because it requires no external AI coding agent:

# In a CI pipeline or script
genty call \
  --harness internal \
  --process .a5c/processes/lint-and-test.js#process \
  --workspace . \
  --no-interactive \
  --json

It executes processes using the SDK's built-in engine, supports all effect types (tasks, breakpoints, sleeps, parallel dispatch), and produces the same event-sourced journal as any other harness.

Package Boundaries

The repository root package.json is workspace metadata for this monorepo. The public packages users install are the scoped packages above.

Monorepo Structure

packages/
  genty/                   # genty agent stack
    cli/                  #   @a5c-ai/genty — CLI binary "genty"
    core/                 #   @a5c-ai/genty-core — agent-core runtime
    platform/             #   @a5c-ai/genty-platform — orchestration platform
    runtime/              #   @a5c-ai/genty-runtime — agent runtime
    ui/ webui/ tui/       #   UI surfaces (web console, terminal UI)
  adapters/               # Adapter family
    sdk/                  #   @a5c-ai/adapters — root SDK + CLI "adapters"
    codecs/               #   @a5c-ai/adapters-codecs — harness codec impls
    cli/                  #   @a5c-ai/adapters-cli
    gateway/              #   @a5c-ai/adapters-gateway
    hooks/                #   hooks-adapter-* (per-harness hook adapters)
    core/ transport/ ...  #   comm-adapter, transport-adapter, etc.
  atlas/                  # @a5c-ai/atlas — knowledge graph + catalog
  sdk/                    # @a5c-ai/babysitter-sdk — core SDK
  kradle/                 # @a5c-ai/kradle — Kubernetes-native forge
    core/ sdk/ cli/ web/
plugins/                  # Installable plugin packages
  babysitter-unified/     #   Unified plugin for all harnesses
blueprints/               # Blueprint marketplace registry
  a5c/marketplace/
library/                  # Process library (methodologies + specializations)

How It Works

+=============================================================================+
|                         /babysitter:call                                    |
+=============================================================================+
|                                                                             |
|   YOUR PROCESS (JavaScript)                   This is the AUTHORITY         |
|   +----------------------------------------+                                |
|   | async function process(inputs, ctx) {  |  Real code, not config.       |
|   |                                        |  The orchestrator can ONLY    |
|   |   await ctx.task(plan, { ... });       |  do what this code permits.   |
|   |                                        |                                |
|   |   await ctx.breakpoint({               |  Breakpoints = human gates    |
|   |     question: 'Approve plan?'          |  (enforced, not optional)     |
|   |   });                                  |                                |
|   |                                        |                                |
|   |   await ctx.task(implement, { ... });  |  Tasks = executable work      |
|   |                                        |                                |
|   |   const score = await ctx.task(verify);|  Quality gates = code logic   |
|   |   if (score < 80)                      |  (not config, real checks)    |
|   |     await ctx.task(refine, { ... });   |                                |
|   | }                                      |                                |
|   +-------------------+--------------------+                                |
|                       |                                                     |
|                       | governs                                             |
|                       v                                                     |
|   +---------------------------------------------------------------------+   |
|   |                      ENFORCEMENT MECHANISM                          |   |
|   |                                                                     |   |
|   |   +-------------+     +------------------+     +-----------------+  |   |
|   |   | MANDATORY   |---->| PROCESS CHECK    |---->| DECISION        |  |   |
|   |   | STOP        |     | What does the    |     |                 |  |   |
|   |   | (enforced   |     | process permit   |     | Permitted: next |  |   |
|   |   |  by hook)   |     | next?            |     | task assigned   |  |   |
|   |   +-------------+     +------------------+     |                 |  |   |
|   |                              |                 | Blocked: halt   |  |   |
|   |                              v                 | until gate      |  |   |
|   |                       +--------------+        | passes          |  |   |
|   |                       | Gate/task    |        +-----------------+  |   |
|   |                       | from code    |                              |   |
|   |                       +--------------+                              |   |
|   +---------------------------------------------------------------------+   |
|                       |                                                     |
|                       | records every decision                              |
|                       v                                                     |
|   +---------------------------------------------------------------------+   |
|   |   JOURNAL: Every task, gate, decision - immutable, replayable       |   |
|   +---------------------------------------------------------------------+   |
|                                                                             |
+=============================================================================+

The difference from simple iteration:

• Process as Code: Your workflow is JavaScript - the orchestrator can ONLY do what this code permits
• Mandatory Stop: Claude cannot "keep running" - every step ends with a forced stop, then the process decides what's next
• Enforcement, not Assistance: Gates block progression until satisfied - they're not suggestions
• Event-Sourced Journal: All run state in ~/.a5c/runs/ by default, with repo-local .a5c/runs/ compatibility reads - deterministic replay and resume from any point

Why Babysitter?

Key differentiators: Process enforcement, deterministic replay, human-in-the-loop breakpoints, parallel execution, and quality convergence (one gate type among several).

Blueprints

Babysitter has its own blueprint system -- and it works differently from what you might expect. A blueprint is not a code module with extension points. It's a set of natural language instructions (markdown files) or deterministic coded processes (JS files) that an AI agent reads and executes. The SDK stores, versions, and distributes the instructions. The AI agent is the runtime.

This means a blueprint can do anything an AI agent can do: install npm packages, generate CI/CD pipelines, set up git hooks, create Terraform configs, modify your linter rules, copy babysitter processes into your project, and interview you about your preferences along the way.

The official marketplace includes blueprints for security (gitleaks, ESLint security rules, audit processes), testing (Vitest/Playwright/pytest setup, coverage gates, TDD processes), deployment (Terraform, Helm, Dockerfiles, multi-environment pipelines), themes (sound effects, design systems, conversational personality), CI/CD (GitHub Actions workflows), and rate limiting (exponential backoff hooks).

To manage blueprints, use the /babysitter:blueprints command inside your harness (or babysitter blueprints:* from the CLI). The agent reads the blueprint's install instructions, interviews you, analyzes your project, and executes the setup -- all within a babysitter orchestration run.

See the full Blueprints documentation for details on how installs work, the marketplace format, creating your own blueprints, and the migration system. Agent harness plugins are covered separately in Plugins documentation.

Compression

Babysitter includes a 4-layer token compression subsystem (built into packages/babysitter-sdk/) that reduces context window usage by 50-67% on real sessions while maintaining 99% fact retention.

All compression hooks are automatically registered by the babysitter plugin -- no manual settings.json configuration needed. Install the plugin and compression is active.

How It Works

Quick Toggle

# Disable all compression
export BABYSITTER_COMPRESSION_ENABLED=false

# Disable a single layer
babysitter compression:toggle sdkContextHook off

# Show current effective config
babysitter compression:config

Config File

Edit .a5c/compression.config.json to persist settings (env vars always take priority):

{
  "enabled": true,
  "layers": {
    "userPromptHook":    { "enabled": true, "threshold": 500, "keepRatio": 0.78 },
    "commandOutputHook": { "enabled": true, "excludeCommands": ["jq", "curl", "docker"] },
    "sdkContextHook":    { "enabled": true, "targetReduction": 0.15, "minCompressionTokens": 150 },
    "processLibraryCache": { "enabled": true, "targetReduction": 0.35, "ttlHours": 24 }
  }
}

Toggle any layer with babysitter compression:toggle <layer> <on|off> or set individual values with babysitter compression:set <key> <value>.

Documentation

Package and Plugin Surface Map

• Package and Plugin Docs Map - canonical public/internal status, docs entrypoints, and coverage notes for active packages and plugins

Getting Started

• Quickstart Guide
• Beginner Tutorial: REST API
• Best Practices
• Migration Guide (0.0.x → v6)

Features

• Adapters - Harness-agnostic runtime across the 12 supported harnesses
• Harness Install Matrix - Per-harness install instructions and invocation tokens
• Process Library - 2,000+ pre-built processes
• Process Definitions
• Quality Convergence
• Run Resumption
• Journal System
• Best Practices
• Architecture Overview

Reference

• FAQ
• Troubleshooting
• Security
• CLI Reference - The core babysitter CLI
• Adapters CLI Reference - The host-side adapters CLI
• Atlas Knowledge Graph

Contributing

We welcome contributions! Here's how you can help:

• Report bugs: GitHub Issues
• Suggest features: Share your ideas for improvements
• Submit pull requests: Fix bugs or add features
• Improve documentation: Help make docs clearer
• Check workspace coverage: docs/workspace-validation.md

See CONTRIBUTING.md for detailed guidelines.

Community and Support

• Discord: Join our community (GitHub invite link)
• GitHub Issues: Report bugs or request features
• GitHub Discussions: Ask questions and share ideas
• npm: @a5c-ai/babysitter

Community Tools

Star History

Contributors

License

This project is licensed under the MIT License. See LICENSE.md for details.