Claude's Workbench – Claude architects and reviews
Claude's Workbench pairs Claude Code with local models like Qwen 27B and MiMo 9B, slicing coding tasks into 150-line units to cut cloud token consumption.
By Dillip Chowdary • Nov 17, 2026 • Source: github.com

At a glance
- Lead Model
- Claude Code via subscription or API
- Local Runner
- OpenCode with Ollama model server
- Supported Systems
- Linux, macOS, and Windows with WSL2
- Python Requirement
- Python 3.11 or higher
- Target Models
- MiMo 9B for chores and Qwen 27B for features
- Task Sizing
- Two to three files and 150 lines on 32K context
In this post
Independent developer a-schaefers released Claude's Workbench, an open-source orchestration tool that pairs Anthropic's frontier model with consumer GPUs to handle software engineering workloads. According to github.com's report, the system positions Claude as a central architect and reviewer while local language models hosted on personal hardware execute code generation and run tests. The architecture delegates repetitive token-heavy programming tasks to domestic machines to reduce cloud subscription and API expenses.
This report details the architectural partitioning, automated skill management, and security boundaries that govern cooperative local and cloud model execution. The analysis is intended for software engineers and systems operators seeking to deploy hybrid local intelligence pipelines without relinquishing code quality or repository integrity. Readers will examine installation prerequisites, command interfaces, and the exact coordination mechanics that keep local agents constrained to isolated workspaces.
01Task slicing and context window limits
Local coding models frequently struggle when assigned broad objectives across multiple project directories, resulting in circular reasoning or abandoned context. Claude's Workbench addresses this degradation by breaking broad architectural designs into bounded, single-concern assignments tailored to hardware capacity. A developer step that failed to complete after an hour as a monolithic instruction landed in thirty minutes when partitioned into three sequential subtasks.
Task dimensions scale dynamically based on the memory and context window supported by the host machine. On standard setups serving 32K context windows, Claude assigns tasks spanning roughly two to three files and approximately 150 lines of code. As local agents complete each step, system telemetry adjusts upcoming assignment sizes, routing subsequent edits in the same directory back to the same agent.
02Self-improving skills and the Hermes feedback loop
The platform captures operational lessons whenever a delegated assignment concludes, recording agent roadblocks, inquiries, and corrective instructions issued by the supervising model. Claude analyzes these aggregated corrections, patches confusing phrasing within existing guidelines, or compiles recurring solutions into fresh operational skills. Agents receive only the specific capabilities pertinent to their immediate task and release those prompt additions once the unit of work finishes.
Borrowing design principles from Nous Research's self-improving Hermes Agent, the framework retains an MIT-licensed security scanner and an outcome-driven verification record. Every prospective skill must pass automated security audits and manual model review before distribution to prevent operational hazards. The system continually evaluates skill utility by contrasting task success rates between executions that loaded the instructions against those that operated without them.
03Model assignment and the frontier review standard
Workloads are routed across available local neural networks according to operational complexity. Routine chores such as renaming variables, adjusting configuration files, or authoring basic unit tests pass to lightweight models like MiMo 9B. Demanding feature implementations, complex refactoring, and subtle defect investigations route to higher-capacity options like Qwen 27B, with the system learning from which local outputs receive final approval.
Advertisement
Tech Pulse Daily
Get tomorrow's pulse first
Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.
The supervising cloud model examines every proposed code diff directly rather than trusting self-reported completion statuses from local workers. Claude rejects work that fails to meet professional standards, refusing to accept outputs merely because they are "good enough for a local model." When unit tests fail, error traces cycle locally back to the coding model, expending expensive frontier tokens only on verified pull requests or stalled sessions.
04Git write protection and isolated worktrees
Security boundaries prevent local coding models from altering central repository state or modifying infrastructure deployment definitions. Claude retains exclusive authority over all version control writes, including branch generation, commit creation, merge operations, pull requests, and continuous delivery setups. Local agents are strictly denied write access to foundational governance files like CLAUDE.md or AGENTS.md, ensuring that project operating constraints remain uncompromised.
Every local coding agent operates within an isolated git worktree alongside dedicated temporary directories and build caches. The primary working checkout remains untouched throughout the generation cycle, blocking unauthorized commands or file modifications situated outside assigned task boundaries. Once Claude verifies diff accuracy and test results, the orchestrator records approved commits under the developer's name without appending artificial attribution lines.
05Installation requirements and runtime terminal commands
Deploying the workbench requires Claude Code with an active subscription or API credentials, OpenCode for agent runtime management, and a local server like Ollama. Host environments must run Linux, macOS, or Windows via WSL2 alongside Python 3.11 or newer and git. The utility inspects host hardware during initial startup to configure compatible model weights and periodically rechecks machine capabilities every week.
- Clone repository Download the source code from the remote repository into a local workbench directory.
- Establish symlinks Link the wb executable into local bin and map the delegate skill into the Claude directory.
- Register plugin Add the cloned repository to the local plugin marketplace and install the workbench package.
- Verify environment Execute the wb doctor command in the terminal to validate all local dependencies and path bindings.
| Command | Action |
|---|---|
| /delegate [task] | Opens the workbench pane and redirects code generation to local hardware |
| /workbench | Displays the real-time activity stream and task state of local agents |
| /delegate-models | Identifies the optimal local language models supported by host hardware |
| /delegate-quit | Halts active local workers while preserving pending work for later |
| /delegate-quit-and-hand-back-claude | Terminates local execution and instructs Claude to complete unfinished tasks |
Note: Native Windows systems require WSL2 execution because the orchestrator depends on POSIX file locks and signal handling that base Windows environments do not support. Developers running local models must also account for continuous electricity costs because heavy GPU utilization during extended builds is not free.
For remote administration away from the development workstation, the wb chat serve command initializes an authenticated local web portal connected to the active Claude process. Developers query connection parameters via wb chat url and can persist the interface as a user-level background systemd service. This background mechanism enables ongoing supervision of long-running compilation and testing routines from mobile or secondary devices.
Frequently asked questions
What hardware and software does Claude's Workbench require?
The system requires Claude Code with an active subscription or API key, the OpenCode agent runtime, an Ollama local model server, git, and Python 3.11 or newer. Supported operating systems include Linux, macOS, and Windows running through WSL2.
How does Claude's Workbench handle local model context limits?
Claude breaks projects into single-concern tasks containing approximately two to three files and 150 lines of code for standard 32K contexts. Task sizing scales dynamically with local hardware and adjusts based on completed steps.
Can local coding models push commits or alter repository history?
No, local agents are restricted to isolated git worktrees and cannot execute git writes, edit project guidelines, or touch continuous integration pipelines. Claude retains exclusive authority over all commits, merges, pushes, and pull requests.
Which local models are recommended for common programming tasks?
Claude assigns lightweight chores like configuration updates and small tests to models like MiMo 9B. Complex feature implementations, architectural refactors, and difficult bug investigations route to larger models such as Qwen 27B.
Sources
Author
Dillip Chowdary
Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.
Related on Tech Bytes
Advertisement