AI Baton

A conductor for your AI coding agents. Claude Code and Codex take turns on one project without losing context, and one agent can review the other's work.

Preview. Works with the Claude Code and Codex extensions, on macOS, Linux and Windows (Windows is new: tested automatically, less used so far). Get it from the VS Code Marketplace, or from Open VSX in Cursor, Windsurf and VSCodium.

Claude adds a function; a new Codex chat already knows what changed and why; then Codex reviews Claude's work and the review can go back to Claude's conversation

The AI Baton panel: Claude and Codex with their latest checkpoints, and Recent work

What it does

Shared memory between agents. When an agent finishes a turn, AI Baton saves a checkpoint: a summary and the agent's last reply. When another agent starts a chat, it gets the other agents' latest checkpoints, so you do not retell what happened. Checkpoints stay in your project (.agent-memory/, ignored by git).
Project knowledge in plain files. AGENTS.md holds one set of rules for every agent; docs/AGENT_HANDOFF.md, docs/STATUS.md and docs/knowledge/DECISIONS.md hold the current state, progress and decisions. Agents read them when they start, including the same agent in a new chat.
Ask for review. Pick the conversation that made the changes and another agent; it critiques that work read-only (the changes, with or without git; or the conversation's replies when no files changed) and gives at least three ranked points with fixes. You can send the review back to the author's conversation (it is copied; you paste it). The author is asked to say what it accepts and to wait for your "ok" before changing code. Agents also run reviews themselves when you say "ask Codex".
Recent work. Sessions that saved AI Baton checkpoints, newest first, with their summaries; reopen a conversation or start a review from one.
Usage and limits. Tokens per day per agent, for this project or all, and the plan limits each agent last reported.
Agents at a glance. Model and effort for new conversations, which agent is working, and a warning when memory hooks are missing or not yet approved.

Ask for review

Pick whose work to review, then the reviewer:

Ask for review, step 1: pick the conversation that made the changes

The review opens beside your code, and you can send it back to the author's conversation:

A review by Codex of Claude's change, with Send to Claude

Usage and limits

Tokens per day per agent, and the plan limits each agent last reported

Getting started

Install AI Baton and at least one of Claude Code or Codex.
Open a project folder and click the baton icon in the activity bar. If VS Code opened the folder in Restricted Mode, trust it first: AI Baton runs local commands, so it stays off (no icon) in untrusted folders.
Click Initialize project. AI Baton offers to install its memory hooks once for all projects. In the project it creates docs/ and .baton.json, creates AGENTS.md and CLAUDE.md or adds an AI Baton section to existing ones (your text stays), and adds .agent-memory/ to .gitignore. Without git, it also takes a first private snapshot of the files in .agent-memory/, so reviews can show what an agent changed.
Codex only: approve the hooks once. In the Codex chat box, click the hooks button next to the model picker ("Review hooks"), then Allow all. The panel shows a reminder until you do.

Then start a new chat in either agent.

Requirements

macOS, Linux or Windows (native Windows, not WSL).
Python 3.9 or newer as python3 on macOS and Linux (on macOS: xcode-select --install); on Windows Python 3.11 or newer from python.org (the py launcher or python). Or set aiBaton.python.
The Claude Code and/or Codex VS Code extension, signed in.
For Ask for review: a git repository or a project set up with AI Baton (without git, the changes come from AI Baton's local snapshots), and the reviewing agent's command-line tool works with your account. A review uses that agent's plan.
Other agents (Copilot, Cline, Continue, ...) can be added to the panel. They do not save checkpoints yet; when you open one, AI Baton copies a request to load the shared context, which you paste into its chat.

What AI Baton stores, reads and writes

The memory hooks run locally: they never use the network or call a model. AI Baton itself sends no telemetry. The agents you use have their own network and data handling.

In your project

.agent-memory/: each checkpoint holds a summary and the agent's last reply (up to 12,000 characters, with obvious secrets such as API keys masked; this is best effort, not a guarantee), older checkpoints in a compressed archive, and which agent is working. Outside a git repository it also keeps local snapshots of the project files in .agent-memory/shadow.git to point at what changed. These checkpoints are given to the next agents, which send them to their providers like any other context. Delete the folder to remove all of it.
.baton.json, docs/, AGENTS.md, CLAUDE.md, .gitignore, as described in Getting started.
Project hook files (.claude/settings.json, .codex/hooks.json), only if an agent has no hooks installed for all projects.

Automatically, when the extension starts

~/.ai-baton/: the bundled command-line tool. ~/.cache/ai-baton/: cached results.

After you agree in the panel

Memory hooks in ~/.claude/settings.json and ~/.codex/hooks.json. They run ~/.ai-baton/bin/baton and do nothing outside projects set up for AI Baton.
Model and effort for new conversations: Claude in ~/.claude/settings.json, Codex in ~/.codex/config.toml.
Agents you add or hide: ~/.config/ai-baton/agents.json.

Read only

Claude Code and Codex logs and transcripts, for token usage, session information and which program a conversation runs in: ~/.claude/projects/, ~/.codex/sessions/.
Plan limits and models: ~/.claude.json, ~/.claude/settings.json, ~/.codex/config.toml, ~/.codex/models_cache.json.

Network

Find agents reads the public VS Code Marketplace listing.
Ask for review runs the reviewing agent's command-line tool, which uses that agent's account and plan.

Removing AI Baton

Uninstalling the extension does not remove the hooks it put into the agents' settings. Before uninstalling, run in a terminal:

python3 ~/.ai-baton/bin/baton uninstall-hooks

Add --project <folder> to also remove hooks from a project's own files. Then delete ~/.ai-baton, ~/.cache/ai-baton and ~/.config/ai-baton. In each project you can delete .agent-memory/ and .baton.json, and remove the .agent-memory/ line from .gitignore. AGENTS.md, CLAUDE.md and docs/ may hold your own work: remove only the part between  and  and the @AGENTS.md lines AI Baton added, and delete these files only if AI Baton created them and you no longer need them.

Remove on an agent's card only hides the agent in the panel; its hooks stay and its checkpoints still reach other agents.

Good to know

AI Baton relies on details of Claude Code and Codex that are not public APIs: log and transcript formats, the cached plan usage, the links that reopen a conversation, and how Codex reports hook approval. An update to either agent can break a feature until AI Baton catches up. The rest keeps working.

Commands

AI Baton: Switch Agent (Cmd+Alt+A / Ctrl+Alt+A)
AI Baton: Ask Another Agent for a Review
AI Baton: Show Recent Work
AI Baton: Initialize Project, Add Agent, Remove Agent, Add Decision, Open Handoff, Copy Handoff Request

Settings

aiBaton.python: the Python interpreter (default python3).
aiBaton.home: leave empty to use the bundled command-line tool; set a folder to use a development checkout.
aiBaton.usageDays, aiBaton.refreshSeconds: usage chart range and refresh interval.

License

Source-available under the PolyForm Shield License 1.0.0. Use is subject to its noncompete and notice conditions; see the license for details. Source and issues: github.com/wwijatkowski/ai-baton.

AI Baton: Claude Code & Codex memory, handoff, review

Wiktor Wijatkowski