agents-workbook
A local proxy that lets Claude Code or Codex display their reasoning on a dashboard, so you can see what the model is actually deciding — and rejecting — while it works.
🔗 Visit agents-workbook
Description
When you hand a coding task to an AI agent and walk away, it's easy to end up trusting a result you never actually watched get built — you see the final diff, not the reasoning, dead ends, or alternatives the model considered along the way. agents-workbook makes that reasoning visible in real time instead of hiding it inside the terminal.
agents-workbook is a local proxy that streams a model's reasoning to a dashboard running at 127.0.0.1:8080 while Claude Code or Codex is working, capturing decision-making, rejected alternatives, and cost analysis per session. It integrates via custom base URLs, tracks CLI sessions separately with output-token counts, supports optional archival through the WORKBOOK_ARCHIVE_PATH environment variable, and deploys locally via Docker or Docker Compose. It's free and open source under Apache 2.0, though because each turn triggers two provider calls instead of one, it does noticeably increase API token consumption.
💬 Our review
The short version: agents-workbook fills a narrow but real gap — auditing an AI coding agent's reasoning in real time, locally, without piping everything into a hosted observability platform.
Hosted LLM observability tools like LangSmith or Helicone capture traces at the API level and are built for production monitoring across many users and sessions; agents-workbook is scoped much narrower, to a single developer watching Claude Code or Codex work on their own machine, with the dashboard staying entirely local. That's a meaningful trade-off: you get simplicity and privacy instead of team dashboards, alerting, or historical analytics across sessions. The real cost to weigh is the doubled API calls per turn — for anyone running long agentic sessions, that's a real token-budget hit, not just a footnote. Worth running when you specifically want to audit whether an agent's stated plan matches what it actually did, or you're debugging why it made a strange decision; not worth the extra token spend for routine coding sessions where you're happy just reviewing the final diff.
💰 Pricing
📊 Global score
🤖 AI-enriched data
Open source, licence Apache 2.0 — gratuit, mais double les appels API par tour (coût indirect en tokens).
Pros
Streaming temps réel du raisonnement vers un dashboard local
Capture les alternatives rejetées et l'analyse de coût
Fonctionne entièrement en local, aucune donnée envoyée à un service tiers
Déploiement simple via Docker / Docker Compose
Gratuit et open source
Cons
Double les appels API par tour → surcoût token significatif
Pas de dashboard d'équipe ni d'historique cross-sessions comme les plateformes hébergées
Scope volontairement limité à un développeur solo sur sa machine