How it works

Compose the work.
Keep the boundaries clear.

Define a pipeline, pick compatible implementations and run it through the approvals and systems you already have.

Define the process

Connect inputs, processing and decisions through a pipeline.

Run in your environment

Use the backends you choose, with limited connections and permissions.

Inspect the evidence

Follow run status, outputs and recorded decisions.

Technical architecture

Separate the business process
from its implementation.

Install a domain's workflows, pick compatible implementations and set up your environment. Business rules stay separate from vendor connections.

01 / DEFINE

Domain pack

The pipelines, mappings, policies and domain inputs that describe your process.

02 / AGREE

Shared contracts

Inputs, settings, outputs and reports mean the same thing whichever implementation runs.

03 / IMPLEMENT

Operators

Installable code that connects models, data systems, tools and compute services.

04 / CONNECT

Deployment settings

Which implementation runs each step in your environment, and how it connects.

Cairn control plane · Checks, approvals, run tracking and recordsExisting runtimes and backends execute the work.
Integrations

Work with your existing stack.

Cairn decides when a step may run and records what happened. Your tools do the work. The line between the two is written down for every integration.

One run, end to end

How a fine-tuning run moves through a typical stack. Cairn runs the gates and keeps the record; each tool does its own job.

1Dataset pinnedAmazon S3
gatereadiness check
2Training jobMetaflow, your GPUs
gatecheckpoint scores
3RegisteredMLflow, after a person approves
4ServedvLLM endpoint, load-tested

Every step, score, cost and approval is kept in the run record.

Run on

Cairn submits, polls and cancels through one engine protocol. Your engine schedules, retries and holds the data.

  • Metaflow
  • Prefect
  • SkyPilot
  • Kubernetes
  • Your own boxes (SSH)

Rent GPUs from

Cairn prices the run first, acquires with a lease and releases on cancel. The provider owns the machine.

  • RunPod
  • Vast.ai
  • Modal
  • Kaggle

Track, register, serve

Cairn decides when to register or promote, after the human gate. Your tracker and server own the versions and the process.

  • MLflow
  • Weights & Biases
  • Hugging Face Hub
  • vLLM

Store in

Cairn owns the store contracts and the artifact relay. You own the bucket and the database server.

  • PostgreSQL
  • SQLite
  • Amazon S3
  • MinIO
  • Google Cloud Storage
  • SQL sources

Call models through one gateway

Every model call goes through a single gateway with redaction and a budget. The provider owns the API.

  • OpenAI
  • Anthropic
  • Ollama
  • LiteLLM

Run durably, observe, notify, keep secrets

Cairn owns which runtime runs agentic graphs, the queries it issues and what it posts. Each tool owns its history, data and vault.

  • Temporal
  • DBOS
  • ClickHouse
  • Slack
  • GitHub
  • Infisical

Integration depth varies: some paths are production-proven, some are integration-tested, and recovery guarantees differ by backend. The integration boundaries in the documentation list each one.

Read the documentation ↗
Questions

Answers before the call.

What platform and ML teams ask first. If yours isn't here, ask it on the call.

Connecting it to your infrastructure

Can I drive it from the terminal or CI?

Yes. The cairn CLI declares, runs and watches pipelines: cairn pipeline, cairn flow, cairn connect, cairn deploy, cairn worker. Anything the console does, the CLI does.

Is there an API?

Yes, an HTTP API with an OpenAPI spec. Launch a pipeline (POST /pipelines/launch), inspect runs, approve or advance a gated step, download artifacts. The console is built on the same API.

Does it work with MCP?

Both ways. Cairn runs an MCP server, so Claude, Cursor or your own agents can launch and inspect runs; and Cairn's agentic steps can call your MCP tools.

Can my systems trigger it, and can it notify mine?

Yes. Inbound events and webhooks start pipelines and feed production scores back. Outbound, run events stream to your listener and notifications go to Slack.

Is there an SDK?

A Python SDK composes and launches pipelines with the same gates and record as the CLI and the API.

Running it

Where does it run?

On your infrastructure. Docker Compose for a single box, a Helm chart for Kubernetes. SQLite to start, PostgreSQL in production.

Where does my data go?

Nowhere new. Data stays in your storage (S3, MinIO, GCS, your databases) and jobs run on your engines. Cairn keeps the record: what ran, what it cost, who approved.

Do I need my own GPUs?

Bring your own, or rent through connected providers (RunPod, Vast.ai, Modal). Either way a run is priced against a budget before it is submitted.

Is it open source?

Apache 2.0 licensed and self-hosted. Early access now; the source opens with the public release.

What happens when a check fails?

The run stops at that gate. Nothing downstream is submitted, the evidence is recorded, and a person decides whether to fix, retry or drop it. A model is never promoted past a failed evaluation.

How does it control cost?

Every pipeline carries a budget. Jobs are priced before they start, refused above the limit, and the actual spend is recorded next to the estimate.

Who approves, and can I enforce that?

Pipelines name their approvers. A gated step waits for one of them, and the approval is recorded with the run. There is no bypass in the pipeline.

Is this a replacement for Airflow, Prefect or MLflow?

No. Those keep doing their jobs. Cairn sits above them, decides when a step may run, and keeps the record across all of them.

Can I write my own pipeline?

Yes, in one YAML file that composes installed steps. It gets the same budget, gates, approvals and record. Operators are installable Python packages.

How do you handle LLM providers and secrets?

One gateway for every model call, with redaction and a budget; providers include OpenAI, Anthropic, Ollama and LiteLLM. Secrets come from Infisical or the environment, never from the pipeline file.

Can I get the record out?

Each run has a downloadable bundle: inputs, outputs, artifacts, scores, cost and approvals.

How do I start?

Follow the quickstart to get a control plane with demo data running, or book a call and we set up the first governed run of one pipeline you already own.

Run it on your own infrastructure.

Apache 2.0 licensed, self-hosted. The quickstart gets a control plane with demo data running; a call gets your first pipeline onto it.

Get started →Book a call ↗