docs/user-guide/ecosystem/overview
Babysitter Ecosystem Overview reference
Docs(../index.md) › Ecosystem() › Overview
Continue reading
Nearby pages in the same section.
Babysitter Ecosystem Overview
**The Babysitter repository is a monorepo: one core orchestration engine plus a family of packages that surround it — a catalog, an agent runtime, adapters, a dashboard, and adjacent products. This page explains what each component is and helps you pick the ones you actually need.**
If you only want to run AI coding work under a deterministic, enforced process, you need very little of this — the core SDK and a harness plugin. Everything else exists to extend that core to more surfaces (CI, Kubernetes, browser), more harnesses, richer observability, and a shared knowledge graph.
**Positioning reminder:** Babysitter's value proposition is **deterministic process execution + complex agentic workflows + policy/process adherence (obedience)** — "Enforcement, not Assistance." Every component below serves that thesis. Quality convergence is one capability of the engine, not the product.
---
On this page
---
What the monorepo is
The repository (a5c-ai/babysitter) holds the entire stack under packages/. At its center is the **event-sourced orchestration engine** (@a5c-ai/babysitter-sdk): your workflow is real JavaScript code, the orchestrator can only do what that code permits, every step is recorded in an immutable journal, and a hook-enforced **mandatory stop** after each step decides what the process permits next. That is the obedience mechanism.
Everything else in the monorepo is built around that engine:
- A **knowledge graph / catalog** (atlas) that supplies harness, plugin, and topology metadata as generated data, so adding a harness is a data change rather than a code change.
- An **agent runtime product** (genty) that packages the agent loop, daemon, and platform layers into a single
gentybinary and exposes the headless/internal harness. - An **adapters family** (20 package directories) that multiplexes a single integration definition out to many harnesses, transports, CI triggers, plugin formats, and lifecycle hooks.
- A **real-time dashboard** (observer-dashboard) that streams the journal over SSE.
- Two **adjacent / forward-looking** components: kradle (a Kubernetes-native Git forge, MVP) and kip-sdk (a memory substrate that is **spec/design only — no shipping code**).
---
The component map
| Component | Package(s) | What it is | Maturity | Read |
|---|---|---|---|---|
| **babysitter-sdk** | @a5c-ai/babysitter-sdk | The core event-sourced orchestration engine: defineTask, the effect/journal engine, the built-in Pi execution engine, token compression. Everything else depends on it. | GA | SDK overview |
| **babysitter (metapackage)** | @a5c-ai/babysitter | Convenience install that pulls in the babysitter npm packages and ships the babysitter bin shim. The recommended human-facing install. | GA | SDK overview |
| **adapters** | @a5c-ai/adapters (+ 19 sibling packages) | A **family** of packages that multiplex one integration definition to many targets: triggers (CI), extensions (plugin compile), hooks (mandatory-stop lifecycle), proxy (140+ providers), tasks (durable breakpoints), and more. | GA | Adapters overview · Adapter types |
| **atlas** | @a5c-ai/atlas | The unified knowledge graph / catalog: YAML graph → generated TS types → JSON snapshots consumed at runtime. Single source of truth for harness metadata, discovery, plugin targets, and topology. Ships the atlas CLI. | GA | Atlas overview |
| **genty** | @a5c-ai/genty + -core / -runtime / -platform | The unified agent product (renamed from "tula") that composes the agent stack into the single genty binary: agent loop, daemon/session/cost, harness integration and governance. Surfaces the internal/headless harness. | GA | Genty overview |
| **observer-dashboard** | @a5c-ai/babysitter-observer-dashboard | A Next.js real-time observability dashboard that streams run/journal events over SSE and virtualizes large run histories. | GA | Observer Dashboard overview |
| **kradle** | @a5c-ai/kradle + @a5c-ai/kradle-cli | A Kubernetes-native Git forge runtime (Argo CD GitOps + Gitea hosting) with a per-org dispatchable assistant. | MVP | Kradle overview |
| **kip-sdk** | no published package (spec-only); path packages/kip-sdk | A **spec/design-only** memory substrate (K/I/P): git-substrate, bitemporal, signed-fact property graph. **No code — Markdown spec only.** | Spec only | kip-sdk overview |
The canonical machine-readable map of public/internal packages, apps, and harness plugins lives at Package & Plugin Map.
---
How to choose
Pick by what you are trying to do:
Install the metapackage @a5c-ai/babysitter and your harness plugin. You get the engine + the babysitter CLI. See Installation and Quickstart. You do not need atlas, genty, kradle, or kip directly — the engine pulls in what it needs.
- **"I just want to run AI coding work under an enforced process."**
Use genty (@a5c-ai/genty-platform) for genty call --harness internal, daemon, MCP serving, and the TUI; and the **triggers adapter** (adapter types) as the GitHub Action / CI entrypoint.
- **"I want to run a process headlessly, in CI, or via the internal harness."**
This is the adapters family's job — codecs drive each harness, extensions compile one plugin manifest into per-harness formats, and hooks enforce the mandatory stop across all of them. See the Install Matrix.
- **"I want to run on a harness other than Claude Code."**
Use the **proxy adapter** (Python LiteLLM bridge, 140+ providers) via adapters launch --with-proxy-if-needed. See adapter types.
- **"I want to free a vendor-locked harness from a single backend."**
Use the **tasks adapter** (the Breakpoints Adapter) — durable backends including GitHub Issues. See Breakpoints.
- **"I want durable human-approval breakpoints that survive cold starts / serverless."**
Use the observer dashboard (/babysitter:observe or genty observe).
- **"I want to watch runs in real time."**
That is atlas and its atlas CLI.
- **"I want harness/plugin/discovery metadata or a queryable catalog."**
Look at kradle — but note it is an **MVP**.
- **"I want a self-hosted, Kubernetes-native Git forge with built-in agent dispatch."**
Read kip-sdk, and understand it is **spec only — there is no implementation to install.**
- **"I'm researching the memory/context-management design."**
---
Maturity at a glance
Be honest about what is shipping versus what is forward-looking:
| Component | Status | What "status" means here |
|---|---|---|
| babysitter-sdk / babysitter / atlas / genty / observer-dashboard | **GA** (v6.0.0) | Versioned, runtime, generally available. |
| adapters family | **GA** | Versioned and consumed by the runtime. A couple of subdirs are thin (see adapter types for the honest per-package notes). |
| kradle | **MVP** | Described by its own README as an "executable MVP runtime and handoff package." Early; expect rough edges. |
| kip-sdk | **Spec only** | Entirely Markdown — SPEC.md + numbered design docs. No package.json, no src/. Treat any kip "feature" as **design, not shipping code.** |
---
Naming notes
- The agent runtime product is **genty** (binary
genty), backed by@a5c-ai/genty-platform. Some older material (thedocs/v6-announcement.mddated April 2026) refers toagent-platform/babysitter-harness; that naming is stale. The README and package metadata are authoritative — standardize on **genty /@a5c-ai/genty-platform**. - "genty harness" in everyday usage means running processes through the **genty CLI on the SDK's internal Pi engine** (
genty call --harness internal). There is no standalonegentycodec in the codecs package. - The **adapters** family keeps its name precisely because it is not genty-specific — it multiplexes for *all* agents.
---
Next steps
- **Start with the engine:** babysitter-sdk overview
- **See how it all fits together:** Architecture & How It Fits Together
- **Understand the adapter family:** Adapter Types reference
- **Install and run:** Installation → Quickstart