What is cairn

cairn is a self-hosted control plane for the work behind AI models: preparing datasets, fine-tuning, evaluating, serving and batch inference. It decides when each step may run, holds the spend to a budget, pauses for a data check or a person, and keeps a record of what happened. Your engines and providers do the work.

The problem

Work on models spends money and changes things. A fine-tune buys GPU hours, a batch job buys tokens, and a promoted model or a new dataset version is what other people build on next.

So the same questions come up every time. Before a run: what will it cost, and who agreed to it? After a run: what ran, on which data, who approved it, and what did it cost? Without a control plane the answers live in scripts, notebooks and memory.

What cairn does

You describe the work as a pipeline: a YAML file that lists the legs to run, their order, and the gates between them. cairn then does five things.

• One launch door. The console, the CLI and a schedule all start a pipeline through the same code. Policy can add a gate, or refuse the launch, before anything is recorded.
• Every leg is its own run. cairn hands each leg to the engine or operator that does the work, then tracks its status and cost.
• Gates. A gate is answered by data (a predicate over earlier results) or by a person. A waiting gate is stored, so it survives a restart.
• Budgets. Pricing is a step you can gate on before anything spends, and flow legs reserve their estimate against a budget before they start.
• The record. Each pipeline keeps its plan, its inputs, the run behind every leg, and every gate’s verdict and approver. Outputs are registered as content-addressed artifacts.

flowchart TD
  start["Console · CLI · schedule"] --> door["Launch door<br/>policy · budget"]
  door --> pipe["Pipeline"]
  pipe --> legs["Legs: flow · task"]
  pipe --> gates{"Gates: data · person"}
  legs --> work["Your engines and providers"]
  pipe --> record[("Record")]

What cairn does not replace

cairn orchestrates and governs. The work itself runs on tools you choose.

Work	Runs on
Data processing	a dataflow engine: the built-in one, Prefect, Metaflow or SkyPilot
GPU training	a provider operator: Modal, RunPod, Vast.ai or Kaggle (or local CPU)
Model calls	your model API, or any OpenAI-compatible endpoint
Durable execution	the in-process runtime, DBOS or Temporal
State	SQLite or Postgres

Words used in these docs

Word	Meaning
Pipeline	The thing you launch: a YAML file of legs and gates.
Leg	One step of a pipeline. A leg that does work is its own run.
Flow	A component a leg runs: a data or ML stage list, or an agentic markdown flow.
Step	A stage or node inside a flow.
Gate	A leg that stops the pipeline until data or a person says go.
Environment	Where a pipeline runs, and which operator does each kind of work there.
Operator	The code that talks to one tool or provider.
Pack	A directory (or installed package) that holds pipelines, flows and policies.

Next

• Quickstart: run cairn locally, launch a pipeline and approve a gate.
• Pipelines, legs and steps: the model in more detail.
• Install and deploy: run cairn for your team.