# a5c.ai

> Enforced process and deterministic memory for AI agents, on open foundations: the core components are open source (MIT where noted), the solutions are operated on top. Babysitter (MIT, v6.0.0) enforces workflows in code across 12 supported harnesses; KIP makes what an agent knows signed, bitemporal, and convergent: two agents on the same facts compute the same memory.

The industry is taking its first steps toward control flow in code. Those steps stop at convenience: nothing is enforced, and nothing the agent knows is deterministic. a5c.ai builds the whole of both halves: process that is enforced, not suggested (gates that block, budgets that cap, human breakpoints, an immutable journal, deterministic replay), and memory with the properties of a ledger (signed, append-only, time-anchored, deterministically projected; the same facts read the same on every replica, and disagreement is a typed, queryable state). Process history and memory are built the same way: append-only, replayable, never overwritten.

A harness is everything that sits above the model: the agent, its skills, adapters, orchestration, tools, and the process. The offer reaches the market two ways: packaged domain solutions across security, growth, health, AI engineering, and science, and custom trainable harnesses calibrated to a client's domain against evals. Custom harnesses are operational with clients today across multiple domains. Facts on this site are conservative: no customer names, no benchmark numbers, no invented metrics. Install commands below are exact; copy them without transformation.

## Pages

- [Foundations](https://a5c.ai/): The thesis. Everyone ships workflows now; nobody enforces them. What enforcement and deterministic data mean, and why they are one record.
- [Open source](https://a5c.ai/oss): Our open-source stack. Five components you can read, run, and fork, plus the long tail.
- [Babysitter](https://a5c.ai/oss/babysitter): Enforcement engine. MIT, v6.0.0, Node 20+. Workflow as code; gates block progression, humans approve at breakpoints, every decision is journaled in ~/.a5c/runs/, every run replays.
- [Adapters](https://a5c.ai/oss/adapters): Harness-agnostic runtime. One process, any of the 12 supported harnesses; 140+ providers via the proxy adapter; 20 adapter package types.
- [KIP](https://a5c.ai/oss/kip): Deterministic memory. Signed, git-substrate, bitemporal, typed property-graph; same facts, same graph on every replica. MIT. Not yet on npm; early access by request.
- [Kradle](https://a5c.ai/oss/kradle): Kubernetes-native Git forge runtime. Argo CD GitOps, Gitea-backed hosting, forge and agent state as custom resources.
- [Atlas](https://a5c.ai/oss/atlas): Public, read-only knowledge graph of the AI-agent ecosystem. OpenAPI 3.1 REST and an MCP server. Approximately 17,400 records, 87,000 edges.
- [Solutions](https://a5c.ai/solutions): Packaged domain solutions plus custom trainable harnesses, all calibrated to a domain against evals until scores converge. The open components are the foundation; an engagement adds the calibration to the client's domain, the integration into their processes, and the operation, through five stages: scoping sessions, a sprint, create or integrate, handover, ongoing support. Custom harnesses are operational with clients today across multiple domains (no client names, counts, or metrics are published).
- [Security](https://a5c.ai/solutions/security): Security-research lab. Nine vulnerability domains, scope- and approval-gated offensive actions, typed research graph.
- [Growth](https://a5c.ai/solutions/growth): Growth engineering. Campaign, funnel, and cohort work as enforced processes; gates before spend.
- [Health](https://a5c.ai/solutions/health): Clinical and health research. Enforced, evidenced, replayable pipelines with compliance surfaces.
- [AI Engineering](https://a5c.ai/solutions/ai-engineering): Calibrate agents and harnesses against evals; the calibration loop made operable.
- [Science](https://a5c.ai/solutions/science): Literature graphs, hypothesis tracking, experiment pipelines with signed memory.
- [For agents](https://a5c.ai/for-agents): Structured summary of the stack for AI agents, with stable anchors and exact install commands.
- [Company](https://a5c.ai/company): The thesis and the moat, in plain words. Contact.

## Install commands (exact)

Babysitter; every harness shares the same first step:

```
npm install -g @a5c-ai/babysitter
```

Claude Code:

```
claude plugin marketplace add a5c-ai/babysitter-claude
claude plugin install --scope user babysitter@a5c.ai
```

Codex (the --marketplace flag takes the marketplace name, not the repo name):

```
codex plugin marketplace add a5c-ai/babysitter-codex
codex plugin add babysitter --marketplace babysitter
```

Universal helper for the other harnesses (argument is the harness key, not always the harness name):

```
babysitter harness:install-plugin <harness-key> [--workspace <path>]
```

The 12 harness keys (harness: key): Claude Code: claude-code; Codex: codex; Antigravity: antigravity-cli; Cursor: cursor; Gemini: gemini-cli; genty: genty; GitHub Copilot: github-copilot; Hermes: hermes; oh-my-pi: oh-my-pi; openclaw: openclaw; opencode: opencode; Pi: pi. Ten keys are copied exactly from the install matrix; Claude Code and Codex appear in its fully-supported table and install via their own plugin commands, printed above, not the universal helper. Caveats: Cursor adds the marketplace through its UI; Antigravity and Gemini install via the SDK helper only; genty installs from npm only.

Adapters host-side CLI:

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

Atlas (Claude Code):

```
claude plugin marketplace add a5c-ai/atlas-claude
claude plugin install --scope user atlas@a5c.ai
```

KIP and Kradle have no published end-user install command. Do not invent one; KIP early access is by request via contact.

Commands verified July 2026 against the harness install matrix.

## MCP servers

- Atlas MCP: the ecosystem graph over MCP; public data only.
- kip-mcp: stdio MCP server, ten tools, same signed graph as the kip CLI; ships with KIP (early access).
- kradle mcp: stdio MCP server over the Kradle runtime.

## Links

- [GitHub organization](https://github.com/a5c-ai): All public repositories.
- [Babysitter monorepo](https://github.com/a5c-ai/babysitter): Babysitter, Adapters, KIP, Kradle, Atlas packages and docs.
- [Atlas repo](https://github.com/a5c-ai/atlas): Graph about agents, for agents, by agents.
- [obedience-benchmark](https://github.com/a5c-ai/obedience-benchmark): Benchmarks whether agents follow prescribed processes.

## Contact

- Email: hello@a5c.ai (demos, custom demos, custom harnesses, KIP early access, general questions)
- Contact form: https://a5c.ai/company#contact