What teams ask before connecting.

Punk is the adaptive runtime for production AI agents. It observes model and tool execution, applies policy, finds repeated work, tests reusable execution paths, and explains the route selected for every served response.

An adaptive runtime is the infrastructure layer that turns prior agent execution into evidence for future routing. It observes work, governs actions, identifies repetition, proves alternative paths, and uses only eligible paths while preserving live fallback. Read the category definition.

Punk is for developers, platform teams, and AI operations owners running agents with meaningful production traffic—especially when model spend, repeated tool work, action risk, or weak explainability has become an operating problem.

No. Punk is infrastructure around models. Your configured live provider remains available for novel or ineligible work, while Punk learns which scoped execution can safely use a verified alternative.

Not necessarily. Punk provides compatible gateway surfaces and execution visibility, but its defining role is different: it connects observed production evidence to governed, reusable execution. It can complement the models, frameworks, and operational tools already in your stack.

For compatible model traffic, the first step can be a base URL and API key change plus identity headers. Tool tracing, side-effect declarations, feedback, semantic web sessions, and deeper learning surfaces may require the Punk SDK or additional integration.

Observe mode keeps the configured live provider as the source of the user-facing response while Punk records traces, route alternatives, policy outcomes, patterns, and estimated ghost savings. It does not serve an optimized shortcut.

When evidence and policy allow, Punk can reuse exact or similar responses, cache safe read-only tool work, reuse stable plans, run tested declarative workflows, or substitute a cheaper model. Novel or ineligible work falls back to the live provider.

Proof means bounded evidence for a scoped route: historical replay, comparison against fresh traffic in shadow, policy allowance, side-effect handling, and an operational fallback or rollback path. It does not guarantee that every future input is correct or that every workload is optimizable.

The router is designed to fail open to the configured live provider when an optimized route is not eligible or fails. Policy can still block prohibited work. Every served response receives a route explanation.

Tools are assigned side-effect levels 0 through 4. Undeclared tools default conservatively to level 3, a user-visible write. Replay and shadow suppress side effects or use dry-run behavior, and high-impact live actions can require policy and human approval.

Punk records an append-only trace history and derived operational data such as patterns, evidence, route statistics, policies, and reusable workflows. Exact storage, retention, redaction, export, deletion, encryption, and access controls must be reviewed for the deployment in scope. Review security boundaries.

The default posture is tenant-local optimization without cross-tenant raw-data use. Aggregate or cross-tenant learning should require explicit opt-in and deployment-specific review.

Punk provides OpenAI- and Anthropic-compatible gateway surfaces and supports configured provider families. Compatibility details and advanced capabilities vary; use the integration documentation for the exact client and feature in scope.

Free is $0 with 10,000 monthly gateway runs. Pro is $99 per month with 500,000 monthly runs and higher resource limits. New verified accounts receive a 30-day Pro trial without a credit card. Enterprise pricing is custom. Provider charges are separate. Compare plans.

Choose one repeated workload with reviewable output, a clear owner, representative traffic, and low or controlled side effects. Support classification, internal record summaries, and read-only research plans are stronger starting points than payments or other high-impact writes.