Sandboxed coding-agent workers for AI supervisors over MCP. Delegate tasks from Grok Bot or any MCP client to Kilo CLI or OpenCode running in per-task git worktrees inside bubblewrap, with any OpenAI-
Unofficial community project. Not affiliated with, endorsed by or sponsored by xAI.
Grok Workhorse lets a supervising AI agent (for example Grok Bot) hand coding tasks to sandboxed coding-agent workers on your own Linux machine. The supervisor calls a small MCP interface (delegatetask, waittask, taskresult, ...). A local daemon creates a fresh git worktree for each task and runs a coding-agent CLI (Kilo CLI or OpenCode today, with more adapters on the way) inside a bubblewrap sandbox, using any OpenAI-compatible model you configure. The daemon then runs your tests itself and returns a short structured result plus the diff. Workers never commit, merge or push: you (or your supervisor) review the branch and decide.
- Repo: https://github.com/mrchatam/Grok-workhorse - License: MIT (vendored skills: MIT, see NOTICE) - Status: v0.3.0 in development (latest release v0.1.0), Linux only
Why: to reduce supervisor (for example Grok) usage. The expensive model plans and reviews; smaller models you choose do the bounded coding work, and v0.3 keeps what the supervisor reads and does per task small (one waittask call, a ~0.5-1 KB brief result, automatic fix rounds and reviews on cheap models). See docs/token-savings.md.
- MCP server over stdio, usable from any MCP client (Grok Bot custom connector, Claude Desktop, Cursor, ...). It forwards to one persistent daemon, so many supervisor sessions share a queue. - Pluggable worker backends. One adapter interface drives different coding-agent CLIs, and each profile picks its backend (adapter matrix). - Any OpenAI-compatible provider: NVIDI
From the project README.
Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:
[](https://mcp.liqiwa.com/s/mrchatam--Grok-workhorse.html)
🖱️ Let Factory Droid control your Mac with the Codex app's Computer Use tool. One-command install, no approval forms, works with any Codex install path.
sandbaseai/deepseek-harness-handbookIndependent, source-backed handbook for DeepSeek AI's official DeepSeek Harness (dsh): agents, plugins, security, troubleshooting, and runbooks.
slxca/opencontextA Model Context Protocol server for persistent project-specific AI agent context.
nettanvirdev/inertiainertia-ai: a desktop workspace for AI agents. Chat, built-in tools, MCP and OpenAPI integrations, Docker/Daytona sandboxes with a live desktop, routines and memory. Bring your own model and keys.
HarjjotSinghh/locumDelegate Grok Bot coding work to your own local Claude Code / Codex CLIs via a custom MCP connector
ucsandman/declickCompile an API, an MCP server, a database or a web page into a CLI your agent can run. Ten engines, one output contract, 4.1x less context than raw MCP.
The top new MCP servers of the week, every Monday. No spam, unsubscribe anytime.