Observe
Request, response, token, cost, latency, identity, and route context.
// OpenAI-compatible integration
Make existing model traffic teach the runtime.
Point compatible requests at Punk, keep the configured live provider as your source of truth, and begin seeing cost, latency, repeated patterns, policy decisions, and route explanations.
Compatibility applies to supported model API traffic. Tool execution, application identity, outcomes, and provider-specific features should be verified for your implementation.
// first connection
Begin with a narrow, reversible change.
Create a Punk account
Generate a key and choose the environment where you want to establish a baseline.
Change the compatible base URL
Keep your existing request shape where supported and send a small test workload first.
Confirm identity and traces
Verify app context, messages, model, tokens, cost, latency, and route explanations before expanding traffic.
Stay observe-only
Collect representative evidence before evaluating caches, workflows, substitutions, or any serving change.
// beyond a proxy
The endpoint is the entry, not the category.
Punk uses compatible traffic as input to an adaptive runtime: observed execution becomes governed evidence, candidates can be replayed and shadowed, and only eligible routes can serve with live fallback.
Govern
Policy evaluation and conservative handling for declared agent actions.
Adapt
Evidence-gated caches, plans, declarative workflows, or model routes where justified.
Limitations: exact API coverage varies. Test streaming, structured outputs, tool calls, multimodal inputs, provider-specific options, and error behavior before routing production traffic.
Connect one workload first.
Use a representative, reviewable model task and confirm the trace before increasing scope.