Description
Agents Workbook is a local proxy that sits between a coding agent, Claude Code or Codex, and the underlying model. It adds one extra tool call to every request: a place for the model to think out loud before it answers, writing down the decision it's making, the alternatives it considered, and what each option would have cost. The project's stated goal is to let a developer check whether an agent's stated plan matches what it actually goes on to do.
Read More
The proxy runs locally through Docker Compose and exposes a dashboard at 127.0.0.1:8080, where each note streams in as the model writes it, with separate CLI sessions kept in separate rows. Claude Code points at the proxy by setting ANTHROPIC_BASE_URL to the local address; Codex is configured through a custom model provider. An optional continuity mode keeps earlier notes available across tool calls within the same turn (on by default for Codex, and enabled for Claude Code with an explicit environment flag once it has been checked against the installed Claude Code version), with configurable limits on session count, memory and idle timeout.
The project documents its own costs plainly. Because it makes an extra call to the model for every turn, and up to four when the model keeps writing notes, it roughly doubles input token usage and adds hundreds to thousands of output tokens per turn, and a reply doesn't start streaming until the hidden note call finishes, up to 60 seconds. It also warns that a workbook note is ordinary model output, so a pasted API key or secret could show up in one.
Agents Workbook does not read or expose a model's protected internal reasoning; the notes are produced through an ordinary tool call, the same as any other tool argument a model can write. Nothing is written to disk unless the operator sets an archive path, and the project's documentation explicitly asks users not to use it to collect or train on reasoning traces from Anthropic's or OpenAI's models. It's released under the Apache 2.0 license and states it is not affiliated with either company.
It's a small, single-developer, open-source project aimed at developers who use Claude Code or Codex and want visibility into an agent's reasoning during a coding session, rather than a product for non-technical users.
- CategoryAI Agent Builders
- Pricing ModelOpen source
- DeploymentSelf-hosted
- PlatformsWeb
Agents Workbook's Core Features
Local proxy adds a "think out loud" tool call to every Claude Code / Codex request
Live dashboard (127.0.0.1:8080) streaming each note as it's written
Separate rows per CLI session on the dashboard
Optional cross-request continuity that keeps earlier notes available within a turn
Configurable continuity limits (session count, memory size, idle timeout)
Works with both Claude Code (via ANTHROPIC_BASE_URL) and Codex (via custom provider config)
Runs entirely locally via Docker Compose; nothing written to disk unless configured
Apache 2.0 open-source license
Getting Started with Agents Workbook
Clone the repo: Get the project from GitHub (softcane/agents-workbook).
Start the proxy: Run docker compose up -d --build.
Open the dashboard: Visit 127.0.0.1:8080 and leave the tab open.
Point your CLI at the proxy: Set ANTHROPIC_BASE_URL for Claude Code, or configure Codex's model provider settings.
Work as normal: Use Claude Code or Codex, and watch its reasoning stream to the dashboard.
Enable continuity (optional): Set WORKBOOK_CONTINUITY_CLAUDE_ENABLED=true to keep notes available across tool calls for Claude Code.
Agents Workbook's Use Cases
- Plan-vs-action auditing
- Coding-session review
- Local debugging
