Mint a model lease for MCPJam-hosted inference
How an SDK eval running OUTSIDE the platform (local dev, CI) uses MCPJam-hosted inference on the organization’s credits, with no provider key of its own. Authenticate like any other /api/v1 route (typically an sk_ key).
The lease is strictly narrower than the key used to get it: one organization, one project, one model, one proxy host, about 30 minutes, a per-lease spend and call cap, and revocable by runId.
The {projectId} segment declares which project’s organization pays, and always wins over any projectId in the body. The literal default resolves to the key org’s Default project — the zero-config CI case, and the same resolution the eval-ingest routes use, so a lease and the results it produces land in the same project.
Authorizations
MCPJam API key (sk_…). Create one at Settings → API keys. Guest sessions cannot use the API, and API keys cannot manage other API keys.
Path Parameters
Project id, or the literal default for the key org's Default project.
Body
Canonical model id. Only anthropic/claude-* and openai/gpt-5* are served, and only if the model has a gateway pricing row.
"anthropic/claude-sonnet-4.5"
Caller-chosen id to revoke by. Generated if omitted.
Per-call output ceiling, clamped to the server maximum.
Response
A short-lived lease and the proxy to present it to.
Present it as the x-mcpjam-harness-lease header on requests to proxyBaseUrl. Never as a bearer token, and never sent to any other host.
Which vendor wire format proxyBaseUrl speaks, derived from the model's vendor prefix.
anthropic, openai Base URL for the metered model proxy. Anthropic leases accept POST /v1/messages; OpenAI leases accept POST /v1/responses.
Epoch milliseconds. Mint a new lease before this; the proxy fails an expired one closed.
Pass to the revoke operation to kill this lease early.

