Developer a-schaefers has released Claude's Workbench, an open-source plugin that redefines the hierarchy of AI-assisted coding. The system positions Claude as the architect and reviewer, while delegating the actual code writing to local Large Language Models running on the user's hardware. This hybrid approach aims to slash API costs by shifting the bulk of token generation to local GPUs, reserving expensive frontier model compute for planning, architecture, and final quality assurance.
The Architecture of Delegation
The core philosophy is "Frontier judgment, local horsepower." Workbench breaks projects into small, manageable pieces sized specifically for local models' context windows. For instance, a model with a 32K context is given tasks limited to two or three files, preventing the degradation in quality that often occurs when smaller models attempt large, complex refactors. Claude determines the design, writes precise specs for each step, and then assigns the work to the most suitable local model via OpenCode and Ollama.
Self-Improving Skill Library
Unlike static prompt engineering, Workbench features a dynamic skill library that evolves with every task. When a local agent fails or requires correction, Claude analyzes the lesson and updates a specific skill file or creates a new one. These skills are scoped to the task type and are security-scanned before being added to the library. This creates a feedback loop where the local agents effectively "learn" the project's specific patterns and pitfalls without requiring traditional retraining, improving performance over time.
Strict Control Over Git and Infrastructure
To prevent local models from making catastrophic errors, Workbench enforces strict boundaries on what agents can touch. Local models are restricted to writing code within their own git worktrees and can only use git for reading. Critical actions like commits, merges, pushes, and CI/CD deployments are locked behind Claude's approval. The system ensures that the frontier model reviews every diff before it is committed to the main branch, maintaining a high standard of code quality that local models alone might miss.
Key Takeaways
- Workbench requires Claude Code 2.1.287 or newer and uses OpenCode to manage local agents.
- It dynamically selects models like Qwen 27B for complex tasks and MiMo 9B for simple chores based on hardware capabilities.
- The plugin is MIT licensed and supports Linux, macOS, and Windows via WSL2.
- A live pane allows users to monitor the interaction between Claude (the driver) and the local agent (the worker).
The Bottom Line
This is a pragmatic, if complex, solution to the cost-quality trade-off in AI coding. By treating local LLMs as unskilled labor supervised by a frontier model, Workbench offers a path to cheaper development without sacrificing architectural integrity, provided you have the GPU horsepower to spare.