M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
B

Babysitter MCP Server

Community
Enforce obedience on agentic workforces and manage extremely complex workflows through deterministic, hallucination-free self-orchestration.
How to install ↓ Compare GitHub source repository ↗
Category
Dev Tools #355 of 438
Stars
★ 1.9k Very popular
Transport
stdio (local process)
Runtime
Node.js 20.0.0+
Credentials
No credential needed
License
MIT
Last commit
40FMRS · D

Babysitter is an MIT-licensed orchestration tool whose central claim is enforcement over agent behavior: processes written in JavaScript are the sole authority, every step ends in a mandatory stop, gates block progression, breakpoints require human approval, and all decisions are journaled and replayable. v6 adds an Adapters runtime so the same processes run across the 12 harnesses in its install matrix. The README documents multiple npm packages, marketplace plugin installs, and a built-in internal harness. It suits teams that need strong process control, auditability, and resumability, but it writes to a state directory, acts on the workspace according to each process, and blueprints can perform broad installation actions. MCP-related capability (for example genty start-server --transport stdio) is inferred from the supplied README; the source provides no verifiable MCP tool list or client config example, so those fields are left empty.

Strongest · Documentation 11/20 Weakest · Reliability 6/20

Reliability
6/20
Security and permissions
7/20
Maintenance
9/20
Documentation
11/20
Setup experience
7/20
Why each score
Reliability 6/20
Static review rests only on the root README plus the repo description; the SERVER MANIFEST is not cached, so the declared MCP tool surface cannot be matched against actual behavior, which caps this dimension. The README claims a genty start-server --transport stdio MCP server and a babysitter-mcp-server binary, but never lists that server's tools, parameters, or initialization handshake, so the core MCP path is unverifiable. The project is a large monorepo with many packages (@a5c-ai/babysitter, adapters-cli, genty-platform, SDK, atlas, kradle, kip-sdk) and layered runtimes with differing Node floors (20.0.0+ vs 22.13.0+), raising dependency and startup complexity while leaving error handling and failure diagnosis unverifiable statically. A CI badge pointing at ci.yml is a verifiable sign of automation, but no committed tests covering the MCP server's key paths are shown. Deductions: missing manifest, undisclosed tool list, heavy runtime dependencies, no static evidence of error handling or replay correctness.
Security and permissions 7/20
No red lines triggered: no malware, credential theft, covert exfiltration, or default irreversible delete/pay/deploy/remote-command behavior is evidenced, and install examples contain no real tokens or secrets. However, permission and data-flow disclosure is incomplete. The product's core premise is 'enforcing obedience' via hooks that register inside the host harness and impose mandatory stops and gates on agent behavior, which is a strong intervention into agent execution permissions, yet the README describes it only in marketing terms with no least-privilege design, scope boundaries, data-boundary statement, or confirmation mechanism for dangerous operations. State is written to ~/.a5c/runs/ by default with repo-local .a5c/runs/ compatibility reads, and the event-sourced journal records every task and decision, but there is no discussion of redaction, access control, or retention limits. The internal harness can discover and delegate tasks at runtime to other harness CLIs (Claude Code, Codex, Gemini CLI, etc.), widening the external-process invocation surface with no stated sandbox or allowlist. The compression subsystem processes prompts and shell output, but its data flow is not disclosed. Deduction: confirmation, scoping, and data-flow disclosure all incomplete.
Maintenance 9/20
The MIT license is explicit and consistent with the description, and there are CI badges, Discord, GitHub Issues/Discussions, CONTRIBUTING.md, and npm publishing channels, so maintenance and response channels exist and are discoverable. Governance and versioning gaps remain: maintainer identity is only asserted via the slogan 'Built with Claude by A5C AI' with no named owner or security-response process; the project is at v6 with a '0.0.x to v6' migration guide, indicating large breaking changes; kip-sdk is 0.0.1/private and unpublished, self-described as internal; and 427 open issues against 1783 stars is a heavy defect backlog, suggesting lagging issue response. Scored as active but with governance and versioning gaps.
Documentation 11/20
Documentation volume is substantial: the README offers a multi-harness install matrix, CLI and adapters-CLI references, a package/plugin and monorepo structure map, compression config examples, blueprint and process notes, and links to FAQ, troubleshooting, security, CLI reference, and process-library sub-docs, so a layered structure broadly exists. Deductions: the manifest is not cached, and the MCP server paths (genty start-server / babysitter-mcp-server) never disclose tool parameters, auth, limits, or cost in the README; harness support is mostly labeled Experimental/Beta while the capability claims read as mature; quantitative claims such as '2,000+ pre-built processes' and '50-67% reduction with 99% fact retention' have no verifiable evidence; and the install chapter carries hidden assumptions (manual Claude Cowork click-through, private-repo requirement for GitHub-synced org marketplaces). Usable, but with hidden assumptions and troubleshooting gaps, hence no higher score.
Setup experience 7/20
Install paths are numerous and must be carefully distinguished: npm global packages (@a5c-ai/babysitter, @a5c-ai/adapters-cli, @a5c-ai/genty-platform), per-harness marketplace commands, and babysitter harness:install-plugin. The README itself warns that the two install tracks 'should not be conflated', adding cognitive burden. Runtime prerequisites are heavy: Node 20+ with the host-side adapters CLI separately requiring 22.13.0+ (for the built-in node:sqlite), plus a supported AI coding harness and initial /babysitter:user-install, project-install, and doctor steps. Concrete Claude Code and Codex commands are given, which are real client-config examples and add some operability; but the manifest is missing, MCP client connection config lacks direct examples, and the primary plugin-marketplace path is fragile outside official marketplaces. Judged workable but multi-step, error-prone, and partly manual.

Static review · not runListed 2026-09-10

Read the FMRS scoring method →

Fit and risk

What it can accessReads local filesWrites / deletes local filesRuns commands or codeUses the network

Best for

  • Teams that need enforced process control and an auditable journal
  • Users working across multiple AI coding harnesses who want one process definition
  • Complex agentic workflows that must be replayable, resumable, and gated by humans
  • Engineering teams that want headless orchestration in CI/CD without an external AI agent

Not for

  • Users who only need one-off chat or single-step code generation
  • Environments where installing Node.js or running external CLIs is not possible
  • Teams wanting fully autonomous automation that cannot tolerate mandatory stops and approvals
  • Users who want to consume the kip-sdk package from npm (it is private/0.0.1, unpublished)

Required permissions

  • Write access to the run state directory (default ~/.a5c/runs/, with repo-local .a5c/runs/ compatibility reads)
  • Create or modify files in the project directory, as determined by the process code being run
  • Perform anything an AI agent can do inside a process, including installing npm packages, generating CI/CD configuration, and setting up git hooks
  • Invoke discovered harness CLIs and external commands during process execution
  • Read and modify .a5c/compression.config.json to configure compression layers

Risks and side effects

  • Process code is the authority, so a poorly written process can cause unintended file changes or command execution
  • Quality-gated processes may iterate repeatedly until gates pass, consuming significant tokens and time
  • Plugin installation uses user scope, affecting all sessions for that user
  • Blueprints can install npm packages, modify linter rules, and write CI configuration, so their instructions should be reviewed before approval
  • The README states babysitter harness:install and harness:discover are deprecated, so older commands may stop working

Setup

Before you start

Runtime：Node.js 20.0.0+

Other optional settings (1)
BABYSITTER_COMPRESSION_ENABLED optional Set to false to disable the 4-layer token compression subsystem; optional, compression is on by default once the plugin is installed.
  1. Ensure Node.js >= 20.0.0 (22.x LTS recommended); the host-side adapters CLI requires Node >= 22.13.0.
  2. Install the main CLI: npm install -g @a5c-ai/babysitter.
  3. To drive a harness from your shell, install the host-side CLI: npm install -g @a5c-ai/adapters-cli, then run adapters doctor.
  4. Install the Babysitter plugin for your harness, e.g. Claude Code: claude plugin marketplace add a5c-ai/babysitter-claude, then claude plugin install --scope user [email protected], restart, and verify "babysit" appears via /skills.
  5. Alternatively use the scriptable installer: babysitter harness:install-plugin <harness> (cursor, gemini-cli, github-copilot, pi, hermes, oh-my-pi, opencode, openclaw).
  6. Run one-time setup: /babysitter:user-install, /babysitter:project-install, then verify with /babysitter:doctor.
  7. For runtime orchestration commands, optionally install @a5c-ai/genty-platform and use genty.

Check that it works

After installing the Claude Code plugin and restarting, type /skills and confirm 'babysit' appears; you can also run /babysitter:doctor to confirm the setup is healthy, or 'adapters doctor' after installing the host-side adapters CLI.

Troubleshooting

  1. Run /babysitter:doctor or genty doctor --run-id <runId> to diagnose run health
  2. After installing the Claude Code plugin, restart and check that "babysit" appears via /skills
  3. If the adapters CLI fails, confirm Node is >= 22.13.0
  4. Resume an interrupted run with /babysitter:resume or genty resume --run-id <runId>
  5. If context usage is too high, inspect settings with babysitter compression:config or disable a layer with babysitter compression:toggle <layer> off
  6. Use --harness internal when no external AI agent is available

Things to try

Once connected, you can ask your AI assistant things like:

  • Use the babysitter skill to implement user authentication with TDD
  • /babysitter:call implement user authentication with TDD
  • /babysitter:plan add pagination to the API
  • /babysitter:doctor diagnose run health

Use cases

Implement features with TDD while enforcing quality gates
Run processes headlessly in CI/CD or scripts via the internal harness
Orchestrate multi-agent workflows by delegating subagent tasks across harnesses
Insert human approval breakpoints at critical steps
Resume an interrupted run deterministically from the journal
Use blueprints to set up security, testing, deployment, and CI/CD configuration

Supported clients

Claude Code
Codex
CursorPartial support
Gemini CLIPartial support
GitHub CopilotPartial support

Listed from the project's documentation, not tested by this site.

Overview

Babysitter is an MIT-licensed tool for agentic orchestration whose stated purpose is to enforce obedience on agentic workforces and let them manage extremely complex workflows through deterministic, hallucination-free self-orchestration. Workflows are defined in JavaScript code (a process): the orchestrator can only do what that code permits, every step ends with a mandatory stop, quality gates block progression, breakpoints require human approval, and every decision is recorded in an immutable, deterministically replayable event-sourced journal. As of v6 it is harness-agnostic via its Adapters runtime, so the same processes run across the 12 AI coding harnesses in its install matrix. Install surfaces include the npm packages @a5c-ai/babysitter (main CLI), @a5c-ai/adapters-cli (host-side adapters CLI, Node >= 22.13.0), @a5c-ai/babysitter-sdk (public SDK/library), and @a5c-ai/genty-platform (optional runtime CLI with call, resume, plan, start-server, tui), plus per-harness plugins. The README also documents a built-in internal harness that runs processes programmatically without any external AI coding agent and can delegate tasks to discovered harness CLIs, a 4-layer token compression subsystem, and a blueprint marketplace of natural-language instructions or deterministic JS processes. The repository also contains an adjacent kip-sdk memory substrate package.

Similar servers

Context7 80 · B

Upstash's official server providing up-to-date third-party library docs for AI coding assistants

★ 63.3k · Tools 2 Compare with this →
Compare with other MCP servers

Source revision feb68abe397a Data synced 2026-10-19 Read the FMRS scoring method