Buddies
Buddies brings focused AI code reviews and a small project process manager into
VS Code. Assign different review instructions to named buddies, review the
current Git working-tree diff, ask for help with selected code, and run your
project's development stack from the status bar.
Preview: The initial Marketplace build supports Apple Silicon Macs. It
requires VS Code 1.90 or newer.
Disclaimer: Buddies is an unofficial extension and is not affiliated with
or endorsed by Anthropic. It requires your own Claude Code installation and
subscription. Reviews consume usage from your account.
Requirements
- An Apple Silicon Mac (
darwin-arm64).
- Claude Code installed and
logged in with your own Anthropic account.
- A Git repository opened as a VS Code workspace when reviewing changes.
The extension holds no API key. Reviews consume usage from your Claude account.
Get started
- Install Buddies and open a Git repository in VS Code.
- Select the robot icon in the Activity Bar.
- Add a buddy, choose a model and effort level, and describe what it should
review—for example, security, correctness, or maintainability.
- Select Review on its card. Buddies reviews all staged and unstaged changes
relative to
HEAD and presents a summary and findings.
Review buddies can read project files for context but cannot run shell commands
or modify files.
Project process stack
Select STACK 0/0 in the status bar to add commands. Each entry runs in its own
integrated terminal. Use START ALL and KILL ALL to control the group, or
select one process to open or start its terminal.
Configuration is saved as workbench.config.json in the workspace. You can
also create it manually:
{
"stack": [
{ "name": "app", "command": "npm run dev" },
{ "name": "tests", "command": "npm run test:watch" }
]
}
Each workspace owns its configuration. Buddies never loads commands from another
project or from a global fallback file.
Help with selected code
Select code in an editor, right-click, and choose Buddies: Help with Selection....
The picker can explain, review, or refactor the selection, or answer a custom
question. Results open beside the editor. A refactor is only applied when you
press Apply replacement, and only if the original selection has not changed.
Use the question box at the bottom of the result panel for follow-ups; the
original selection and earlier answers remain in context. Press Cmd/Ctrl+Enter
to send without leaving the keyboard.
Selection help uses Claude Sonnet by default. It can be overridden in
workbench.config.json:
{
"selectionHelp": {
"model": "claude-sonnet-5",
"effort": "medium"
}
}
How a buddy runs
Every buddy runs through the Claude Agent SDK — Claude Code as a library.
It carries your own Claude Code login, so the extension holds no API key and
.env needs no token.
A subscription token sent straight at the Messages API only reaches Haiku; every
other model answers 429. The Agent SDK is the sanctioned client, so it reaches
every model your plan covers. That is why the direct-API path is gone.
The model picker lists what your login can actually run, asked of the agent
itself. default means "pass no model" and lets the agent pick.
A buddy may read the files around a diff — Read, Grep, and Glob are
pre-approved. Bash, Write, and Edit are removed outright: a reviewer must
never change the tree it is judging. Buddies do not load the reviewed project's
CLAUDE.md, so a review is only ever your buddy instructions plus the diff.
To override either picker list, put a claude block in workbench.config.json:
{
"claude": {
"models": ["sonnet", "opus[1m]"],
"efforts": ["default", "high", "xhigh"]
}
}
Check it
For contributors, clone the repository, run npm install, open it in VS Code,
and press F5 to launch an Extension Development Host.
npm run check # offline assertions
npm run check:live # also runs one real review, which costs usage
Troubleshooting
- Claude binary was not found: install Claude Code, then restart VS Code so
it inherits your updated environment.
- Login expired: run
claude in a terminal and sign in again.
- Model refused or usage limit reached: edit the buddy and choose an
available model, or wait for the usage window to reset.
- Nothing to review: commit a baseline first, then make staged or unstaged
changes in the opened Git repository.
See SUPPORT.md for support and PRIVACY.md for how
project data is handled.