Attribute traffic
Separate environments, teams, apps, and pseudonymous subjects instead of treating coding activity as one undifferentiated bill.
// Claude Code integration
See how coding-agent traffic runs.
Route supported Claude Code model traffic through Punk to attribute usage, preserve application identity, apply configured policy, and explain the runtime path—starting without optimization.
Punk is independent of Anthropic. Claude and Claude Code are products of their respective owner.
// useful for platform teams
Make usage and policy visible by workload.
Apply runtime policy
Evaluate supported model activity and instrumented tool actions against organization-specific rules.
Explain routes
Record why the configured live provider, cache, block, or another eligible path was selected.
// do not confuse traffic with actions
A model connection is not a workstation sandbox.
Coding agents can read files, edit code, invoke shells, access services, and create external side effects. Observing model traffic alone does not govern all of that behavior.
Model-path coverage
- Supported requests and responses
- Usage, cost, and latency context
- Application and subject identity
- Configured model policy and route explanation
Requires explicit integration
- Local file and shell actions
- Repository permissions
- Secrets and network access
- Approvals and external writes
- Task success or code quality outcomes
Use conservative scope. Verify current Claude Code configuration support, authentication, streaming, tools, and failure behavior in a non-production environment before broader rollout.
Start with visibility you can verify.
Connect a bounded team or environment, inspect the trace, and add action-level controls only where they are truly instrumented.