One session. Swap everything underneath.
The session is the durable thing — one id, one transcript, one place memory lives. The model + CLI driving it is a block. The compute it runs on is another. Both swap mid-conversation. Run the same intent through several combinations in parallel and curate the winner.
$ curl -X POST https://gateway.cerver.ai/v2/sessions \ -H "Authorization: Bearer $CERVER_API_TOKEN" → { "session_id": "sess_a4f7c9...", "status": "ready" }
Free at hobby scale. Free for the agents you build on it.
That infrastructure starts with the session.
Yes, "session" sounds boring. So does plumbing. Nobody brags about pipes until the kitchen is underwater. The session is where memory, tools, compute, billing, recovery, and user state come together. Once it becomes important, it needs a clean interface your product and agents can depend on.
Cerver is the API for that layer.
Cerver gives you one interface for creating, running, switching, measuring, and closing AI sessions. But a good session layer should not only be reliable.
POST /v2/sessions POST /v2/sessions/:id/run-llm POST /v2/sessions/:id/compute GET /v2/sessions/:id/metrics POST /v2/sessions/:id/close
Reliable is the baseline. Powerful is the point.
Cerver sessions are stable, transparent, and hard to accidentally break. They keep memory, tools, compute, billing, recovery, and user state together while your AI stack changes underneath. And they do more than hold state: they can switch models, move compute, track cost, recover runs, and spawn other sessions.