For developers

The workload catalog,
one import away.

A Python SDK, a TypeScript SDK, a CLI, and an MCP server for agents like Claude Code and Cursor — four ways into the same currently offered chat, vision-language, and fine-tuning catalog.

Read the docs →See pricing

pip install commoncompute·npm i @commoncompute/sdk

quickstart.pypython

1# Submit a chat job with a hard spend cap
2from commoncompute import Client
3
4cc = Client(api_key="cc_live_…")
5
6job = cc.submit(
7    "mlx_llm",
8    {"messages": [{"role": "user", "content": "Hello"}]},
9    data_class="public",
10    marketplace_execution_risk_acknowledged=True,
11    max_spend_usd=50,  # hard cap
12)
13
14print(job.price_usd, job.eta_seconds)  # estimate captured at acceptance
15result = cc.result(job)

Examples

Idiomatic SDK, one-liners per workload.

Chat

Stream a response from a model in the live catalog.

chat.pypython

1response = cc.chat.completions.create(
2    model="qwen3-4b",
3    messages=[{"role":"user","content":"Hello"}],
4    data_class="public",
5    marketplace_execution_risk_acknowledged=True,
6)

Migrations

Already running elsewhere? Swap the endpoint.

The native SDK is the recommended path for new projects. Existing OpenAI-compatible chat clients can point their base URL and API key at Common Compute.

Coming from
What you have
What changes
Coverage

OpenAI

OPENAI_BASE_URL + OPENAI_API_KEY

point both at Common Compute

chat completions

Documented controls

Controls you can inspect before adopting the platform.

Preflight quotes

The API returns the applicable unit price and an estimated maximum before acceptance. Final billing follows measured usage under that quoted rate; estimates are not fixed-price guarantees.

Gated execution

Jobs run in native runners on provider Mac computers, reached only through a signed task assignment and a vetted-input gate. Per-job inputs and outputs are scoped to that job; runner processes and model weights may be reused between jobs for warm-start latency. API request logs keep route, status, latency, client IP and user agent for 30 days — never request or response bodies.

Execution receipts

Completed jobs expose an authenticated receipt containing the execution and billing evidence recorded by the platform. A receipt is accountability evidence, not proof that a provider could not read assigned plaintext.

Idempotent creation

Reuse an idempotency key with the identical request to recover the same resource. Reusing it for different content is rejected; this does not promise exactly-once physical execution after every failure.

Per-job spend limit

Set max_spend_usd to reject a request whose preflight maximum is too high and cap the completion charge. It is not a promise to pause a running model at an exact fractional unit.

Compatible chat surface

The chat endpoint follows the OpenAI Chat Completions shape for the documented subset. Moving between providers still requires model, capability, authentication, and behavior testing.

Walkthrough

Review the first-job flow.

Setup outline

Submit a bounded first job

An illustrated setup sequence. The live catalog and exact-model availability check remain authoritative at run time.

Sign up at commoncompute.ai/signup and add a card.
Copy your API key from the dashboard.
npm install @commoncompute/sdk.
Submit a small chat job with a spend cap.
Inspect the signed receipt the network returns.

Start with one bounded job.

No cluster setup or minimum commitment. Review a quote and the marketplace trust model before submitting work.

Get started →Read the docs