Comparisons
Cerver vs. OpenRouter
A router for requests. A runtime for agents.
"So… OpenRouter?" is the most common reply to the Cerver pitch. Same sentence on the surface — one API, any model. Opposite economics underneath.
OpenRouter — routes requests
metered
Cerver — runs agents
— or bring your API key
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_KEY" \
-d '{
"model": "openai/gpt-5",
"messages": [
{ "role": "user", "content": "Fix the failing test" }
]
}'
# → text describing a fix.
# applying it, running the tests, retrying — your job.
curl -X POST https://gateway.cerver.ai/v2/sessions \
-H "Authorization: Bearer $CERVER_API_TOKEN" \
-d '{
"session_name": "fix-tests",
"harness": "claude",
"compute": { "provider": "local" }
}'
curl -X POST .../v2/sessions/SESSION_ID/run-llm \
-H "Authorization: Bearer $CERVER_API_TOKEN" \
-d '{ "input": "Fix the failing test" }'
# → the agent edits the files, runs the tests,
# and replies when they pass.
A router picks which model answers a request.
A runtime runs the whole agent — on the compute you choose, billed the way you choose.
Stuck mid-task? The agent asks another agent for a second opinion — full context, like a doctor referring a patient.
The loop runs on your machine until the tests actually pass — not a text suggestion you apply yourself.
Start on the laptop, finish on the server. The session and its transcript move; the work doesn't restart.
A runaway overnight agent is one API call away from stopped — or from your keyboard.
The transcript persists. Pick up Friday's session on Monday with everything it knew.
Private repos, internal APIs, VPN-only services — the agent reaches them because it runs where you run.
OpenRouter ranks models by usage share. We publish what models actually cost — list prices plus observed cost per turn from metered runs.
Real tasks with test suites; a model passes only when every test goes green, retries priced in. 11 models, public verdicts, receipts.
Cost-to-done needs per-session metering with verification in the loop. A request proxy can't measure it — a runtime does it as a byproduct.
Not rivals — different layers that compose. A Cerver session can even use OpenRouter as its model backend. The router picks the model; the runtime runs the agent. And the economics are opposite poles: OpenRouter's pitch is "no subscriptions" — Cerver's is that your subscriptions finally do the work. If you pay for Claude Max or Codex today, the cheapest agent fleet you can run is the one that uses them.
See a session run on your own machine.
Start an agent, watch the transcript live, move it to another computer mid-task — and pay nothing per token beyond the subscription you already have. Cerver's fee is a flat $2 per 1M tokens.