cairn

An open-source control plane for governed ModelOps.
Fine-tune, evaluate, promote, and serve models as single governed commands — cost-capped before the GPU spins up, policy-gated, and audited end to end.

Docs · Install · ModelOps lifecycle · Governance · Waitlist

What is cairn?

cairn runs the work behind AI as governed commands. You declare an outcome — gate a dataset, fine-tune a model, evaluate it against a baseline, promote it, serve it — and cairn drives the tools underneath, enforces the guardrails, and records exactly what ran.

The engine is domain-neutral. Everything that knows about a domain lives in a pack; everything that touches a tool is an operator. The engine never imports a pack — that one-way dependency is what lets the same machinery run fine-tuning, dataset checks, model evaluation, or incident response, and lets any operator inherit the budget, policy gate, and audit for free.

cairn is open source (Apache 2.0) and self-hosted by default: your keys, your GPUs, your logs — nothing leaves your perimeter. A managed cloud is in private beta for teams who would rather not run it themselves; the engine is the same either way.

Why governed?

This is what separates cairn from a bare agent loop or a black-box tool — the work runs behind deterministic guardrails you can inspect afterward:

• Hard budget — every run accrues cost in a scope; over the cap it raises BudgetExceeded, so a fine-tune can't quietly overspend.
• Policy + human approval — proposed actions pass declarative policy rules before any executor runs; high-risk ones (promote a model, destroy a resource) pause for a human and resume durably from a checkpoint.
• Reproducible runs — every run stores its validated inputs, a content- addressed schema hash, and the exact operators it resolved to. Replay any run; explain any result.
• Tamper-evident audit — every model call, tool call, and approval lands in a hash-chained, append-only log: what ran, what it cost, who approved it.
• Citation validation — for synthesized answers, every identifier the model cites is exact-string-matched against its context; unmatched citations are stripped.

Quickstart

The engine ships as the cairndev distribution (import name cairn). The cairndev wheel carries the builtin operator packs (14) and the core contracts (13); their sources live under packages/ in this repository. The public release is five wheels: cairndev, cairn-client, the optional Metaflow and SkyPilot engines, and the example pack cairn-pack-hello.

git clone https://github.com/cairndev-sh/cairn && cd cairn
uv sync --extra dev
uv run cairn --help

A domain pack is a data wheel the engine discovers through an entry point; the example pack shows the shape, and a governed run is a declared pipeline:

CAIRN_PACK_PATH=examples/hello-cairn uv run cairn pipeline ls
CAIRN_PACK_PATH=examples/hello-cairn uv run cairn pipeline explain hello-cairn/profile-and-publish

Those are the commands the public CI job runs on every change. The container image is built from Dockerfile.engine, the Helm chart lives under deploy/helm/cairn, and the optional extras are auth, dataflow-engine, dev, hello, infisical, kernel, metaflow, nats, oauth, optuna, pii, rag, rag-pgvector, rag-pinecone, rag-qdrant, rag-sqlite-vec, rerank-cohere, rerank-cross-encoder, runtime-dbos, runtime-postgres, runtime-sqlite, runtime-temporal, skypilot, ssh, storage-postgres, telemetry, train. Full setup: Installation · CLI reference.

The ModelOps lifecycle

Each stage is a normal governed run; chained together they form one auditable spine — as a template's subgraphs, or as an explicit pipeline.

flowchart LR
  D[Gate dataset] --> P[Prep · scrub · split]
  P --> T[Train · local or remote GPU]
  T --> E[Evaluate · deterministic gates]
  E --> H{{Human approval}}
  H --> R[Register · MLflow]
  R --> S[Serve · vLLM]

Deterministic gates decide pass/fail; LLM judges are advisory and off by default. Promotion pauses for a human. Read the walkthrough in the ModelOps docs.

One engine, many surfaces

The same governed core is driven from two places in this repository:

• CLI — the full control surface: run, preflight, replay, approve, deploy. → CLI reference
• HTTP + SSE API — every control-plane action is an endpoint, including live run streaming. → HTTP API

Documentation

Status & license

cairn is in early access and under active development. The API and command surface may still change between releases; see CHANGELOG.md.

Licensed under the Apache License 2.0 — see LICENSE. Contributions are welcome: read CONTRIBUTING.md first. Interested in the managed cloud or a say in what ships next? Join the waitlist.