Our open-source stack.

AI is early, and the standards for agent infrastructure are still forming; only communities and wide adoption will make them hold. So parts of our stack are open. Read them, run them, and taste what the solutions are built on.

Start with Babysitter GitHub

01The components

Process, data, runtime, graph.

Each of these exists because an agent once said done and nothing could prove it. Observability tools measure a run after the fact; this stack refuses the step that would break the process.

The process pillar enforces what happens next. The data pillar makes what is known deterministic. A cluster runtime carries both, and a graph describes the ecosystem they operate in. Throughout, a harness is everything that sits above the model: the coding agent, its skills and tools, the process around them, and its memory; Adapters lists all 12 supported.

Process

Babysitter

MIT

The enforcement engine. Define the workflow in code; gates, budgets, breakpoints, and the journal are enforced at runtime.

In practice: a run cannot skip a gate, and when you ask how it decided, the journal replays the answer.

Adapters

The harness-agnostic runtime every run sits on. A harness is integrated by describing it, so the same gates and journal hold on all 12.

In practice: write the process once; it runs unchanged on Claude Code, Codex, Cursor, and 9 others.

Data

KIP

MITpre-release

Every write is a signed fact on a git substrate; every read is a pure projection of the fact set. Not yet on npm; early access on request.

In practice: two agents on the same facts compute the same memory, byte for byte.

Runtime

Kradle

A Kubernetes-native Git forge runtime. Argo CD GitOps, Gitea-backed hosting, and a per-org dispatchable assistant.

In practice: every run, repo, and approval is a declarative resource your team can list and audit.

Graph

Atlas

Graph about agents, for agents, by agents. A public, read-only knowledge graph of the AI-agent ecosystem, over OpenAPI 3.1 REST and MCP.

In practice: your agent's answers cite records a reviewer can look up.

02Composition

How they compose.

Adapters is the base plane every run sits on. Babysitter enforces the process above it. KIP holds what the run knows: signed facts underneath, one deterministic projection on top. Atlas is the catalog the runtime reads to discover harness capabilities. Kradle is the cluster boundary the whole stack runs inside.

Each piece stands alone. Together they are one record: what happened and what is known, built the same way, so a postmortem starts from a replayable record instead of a memory of one.

03Start

The developer path starts with one install and ends in a terminal.

One install from now, your next run keeps a journal, stops at its gates, and replays on demand.

Start with Babysitter Read about the data pillar

Want a harness we don’t list yet? Tell us.

04The long tail

Smaller tools, same workshop.

Public repositories that grew out of the same work.

a5c-ai/kanban

Git-native, decentralized kanban board with web UI, TUI, CLI, and a VS Code extension.

github.com

a5c-ai/forge

Git-backed, agent-native distributed code forge. Cloning a repo clones its collaboration state.

github.com

a5c-ai/obedience-benchmark

Benchmarks whether agents follow prescribed processes rather than merely producing correct outputs.

github.com

a5c-ai/babysitter-breakpoints

Lightweight breakpoint manager with API, queue worker, and web UI.

github.com

a5c-ai/registry

Agents registry. Installed via the a5c GitHub Action.

github.com

Want any of this operated for your domain? Request a demo.