Give your agent somewhere to think loud watch its decisions unfold live

agents-workbook is a local proxy that adds one tool to every request going out of Claude Code or Codex, giving the model a place to think out loud. The model writes notes that appear on a dashboard at 127.0.0.1:8080, while the client receives the ordinary reply. The default wording asks for the decision the model is making, the alternatives it rejected, and what each would have cost, at whatever length the problem earns. The project was built to answer whether an agent's stated plan matches what it goes on to do.

The proxy runs via Docker and supports Claude Code and Codex with specific environment variables. It includes continuity features: Codex cross-request continuity is enabled by default, while Claude continuity requires explicit enabling. When continuity is enabled, the proxy keeps earlier workbook exchanges in memory until the active user turn reaches a final answer.

The tool has warnings: it burns tokens, with each turn making two provider calls instead of one, and up to four when the model keeps writing notes. The answer starts later because the proxy waits for the full note before the reply begins streaming, with a 60-second timeout. Notes can leak secrets, and the tool does not read protected internal reasoning; notes are ordinary model output. It is not intended for data collection or training.