One endpoint for every agent harness.

Point any harness at api.wave.online. Runtime picks the model, meters every token, keeps the memory, and writes the audit trail — no vendor SDK, no code changes.

Claude CodeOpenAI SDKVercel AI SDKLangChainLlamaIndexCrewAIMastra

POST /v1/route
 -X POST https://api.wave.online/v1/route \
 -H "authorization: Bearer $WAVE_KEY" \
 -d '{ "agent": "support-triage", "policy": "cheapest-that-passes",
  "budget": { "usd": 0.02, "p95_ms": 900 }, "memory": "ses_8f21c4",
  "input": "summarize the ticket thread" }'

200 OK · routed qwen2.5:3b · 31 prompt + 11 completion tokens · metered on the live rail

ROUTE — one call reaches every provider; a policy decides at request time.
METER — tokens, seconds and images priced as they happen, attributed per agent, per customer, per run.
REMEMBER — session and long-term memory travel with the request.
AUDIT — every prompt, route, cost and output written to an immutable trace you can export or scope.

Status: routing and metering are live (api.wave.online/v1/dispatch); the normalized /v1 front-door on this host is activating — honest until it's GA, no fabricated transport.