claude-tap

Why we included this project

When a coding agent does something unexpected, it rarely explains itself. claude-tap runs as a local proxy between the CLI and its model API, and its browser viewer shows you what the model actually saw: system prompts, conversation history, tool schemas and calls, streaming responses, token usage, and the diffs between requests. It works with Claude Code, Codex CLI, Gemini CLI, OpenCode, and several other coding agents, and comparing adjacent requests makes it easy to spot which message or tool parameter change broke a run. Everything stays on your machine, auth headers are redacted, and you can export a single self-contained HTML trace to share with a colleague without uploading anything to a hosted dashboard. Teams that need to inspect, explain, or archive agent behavior will find this a practical debugging aid.