← Back to archive

Repo of the Day

a-schaefers/claudes-workbench

Published: Nov 21, 2026

Open repository ↗

A frontier model leads your local LLMs: Claude architects the project, breaks it into bite-sized tasks, hands each agent the skills it needs, and reviews every diff to its own standard. Your GPU does the heavy lifting. For Claude Code + OpenCode

Summary

Claude's Workbench is a Claude Code plugin that connects Claude Code with OpenCode, so a local LLM running on your GPU can write code inside an ordinary Claude Code session. Claude stays the lead: it splits a task into steps, assigns each step to a model sized for it, reviews what comes back, and sends work back until it is right. The plugin is Python, standard library only, and MIT licensed.

What it is useful for

  • Delegating code writing to a local model. On bigger coding tasks, the README says that once a session is already open, handing the coding off to the local model typically costs Claude about half the tokens of writing the code itself; the repo documents the measurement in THESIS.md.
  • Reusable know-how. Claude keeps and grows a library of skills and hands relevant ones out per task, so local agents work with the same guidance Claude has.
  • Non-code assets. Optionally, Claude directs local image and audio models for icons, illustrations, voice lines, sound effects and music, and checks each output. Small speech models run on the CPU; music wants a GPU.
  • Visibility. The pane's Stats tab shows this machine live — the GPU, the running task, every model in use — plus history such as tasks landed, Claude tokens saved, energy and streaks.
  • Safe review and optional extras. Agents work on their own branches, so your checkout stays untouched until you accept. With probe installed, Agent code navigation lets agents read a file's outline and one function instead of the whole file. Web chat in Settings lets you message Claude and answer its prompts from your phone over your own network, such as Tailscale.

How engineers can use it

You need Claude Code with a subscription or API key, OpenCode 2.3 or newer (the README documents curl -fsSL https://opencode.ai/v2/install | bash, or opencode upgrade --major from 1.x), plus Git 2.34 and Python 3.11 or newer on Linux, macOS or WSL2. A local model is also needed: an existing server that speaks the OpenAI API is enough (tool calls and a 32K context window recommended, 16K at minimum), and if you have none, Claude suggests one that fits your hardware and asks before installing or downloading anything.

Install and set it up:

claude plugin marketplace add a-schaefers/claudes-workbench
claude plugin install workbench@workbench
"${CLAUDE_CONFIG_DIR:-$HOME/.claude}"/plugins/marketplaces/workbench/bin/wb setup

Then, inside your project in Claude Code, run /workbench. The first run completes setup and proves itself on a tiny sample task. /workbench-quit stops the agents and keeps their work for later, /workbench-quit-and-hand-back-claude has Claude finish those tasks itself, and /workbench-update installs the latest version.

One limitation to note: it is tested in the Claude Code CLI only, not the Claude desktop app.