RAPI · Benchmark Beta — warming
Model-vs-model performance per endpoint — published with method once the index is warmed by a multi-agent swarm.
Ecosystem / Intelligence
The swarm remembers.
Every request an agent makes carries a signal — which model asked, which endpoint answered, how
fast, how cleanly. RAPI turns that individual telemetry into a shared performance
index: contribute your model on a request and you get back real intel on how models and
endpoints are actually performing, so your agent can steer its own next call. It is not
priced — it simply asks for the contribution.
X-LLM-Model → X-Tool-Stats (contribution-gated, not priced) Swarm phase Delivery — the scout's report: each request feeds a shared index an agent can read to steer its next call.
Beta
The telemetry engine behind RAPI is live and running. The public index is in beta: today it is warmed almost entirely by our own agent swarm, so the numbers are real but not yet a neutral, multi-vendor benchmark. We publish the method and the provenance alongside it, and we do not dress a single-operator dataset up as a leaderboard.
Contribute to consume
A feedback loop, not a black box.
The mechanic is deliberately simple and symmetric. Send the model you are using on a request (X-LLM-Model), and the response comes back with a compact performance report (X-Tool-Stats) — your model's rank on that endpoint, its response time, cache behaviour, token cost. Every agent that contributes its signal makes the index sharper for the next one. Free-riding earns nothing; contribution earns the intel.
Stated precisely
swarmsDB hands you the intel. Your agent does the steering.
RAPI measures and reports; it does not route. The index tells your agent how models and endpoints are performing — the decision to switch models, retry, or pick a different endpoint stays in your agent's own reasoning loop, on its next call. It is a shared memory the swarm reads from, not a dispatcher that reaches in and reroutes you.
The public benchmark
Coming when the data earns it.
A neutral, model-versus-model, endpoint-by-endpoint leaderboard needs traffic from many independent agents before it means anything. Until the index is warmed by more than our own swarm, this stays an honest empty frame rather than a chart that overstates a single-operator sample.
Measured on our own agent-swarm workload. RAPI's contribution-for-intel loop is live; the public benchmark is held to a higher bar than "it runs" — it ships when the data behind it is neutral enough to defend.
More in this area
Not a framework, a swarmWhy swarmsDB is an ecosystem — three swarm phases, phase for phase.Auger — curationThe framework that turns any API into a shape-typed, canary-guarded source.The data-network commonsA shared cache the swarm warms — and everyone rides. Warm it, pay less.Earned freshnessHow long the commons stays warm — a TTL justified per endpoint, never guessed.Orchestration surfaceOne typed toolset reaches the whole catalogue and drives fan-out and scheduling.
Point one agent at it.
Connect over MCP or REST. If your agents read more than one source, the context bill is the first thing you will see move.