RexyMCP

rexyMCP

Let Claude|Antigravity be the Architect. Let a local model do the execution.

rexyMCP is named after Rexy, my cattle dog. Rexy lives to herd — to keep the flock moving in one direction, to circle back the stragglers, to never let the work wander off. rexyMCP herds your local LLM the same way. Claude Code is the Architect: it decomposes your idea into a product plan with milestones and spec'd phases, then reviews every result. A local model is the Executor: it writes the code, one bounded phase at a time, on an OpenAI-compatible endpoint (vLLM, LM Studio, Ollama). And in between — keeping the Executor on task, in-bounds, and honest, turn after turn — runs rexyMCP: parsing its messy tool calls, catching its loops before they spiral out of control, feeding compiler errors straight back into the next turn, and refusing to let the LLM touch anything outside the repo. The Architect designs; the Executor codes; rexyMCP keeps the herd together and moving toward the goal.

Your Architect runs in either Claude Code or Google Antigravity — rexyMCP ships the same skills and MCP tools to both, so the workflow is identical whichever you drive.

Two things make rexyMCP stand out:

1. It is purpose-built for small local models. A 7B–27B model will make mistakes a frontier model never would. So every layer of the loop has a specific answer: a forgiving tool-call parser that repairs malformed output instead of aborting the turn, a post-edit verifier that feeds compiler diagnostics back for an immediate retry, a governor that catches repetition and runaway output, security-scoped tools that can't escape the target repo, and a read-before-edit rule that stops blind overwrites. Models that would be too unreliable to use unguarded become productive executors for bounded, spec-driven work — because the plumbing corrects their failure modes automatically.

2. It runs a virtuous cycle of continuous improvement. Every dispatch is a labeled data point — gates, parse-failure rate, length-truncation rate, turns, peak context, the Architect's verdict — written to a shared telemetry store. rexymcp runs, rexymcp scorecard, and rexymcp profile turn that passive exhaust into evidence: which model, at which settings, on which kind of work, actually earns its keep on your codebase. And recurring failure patterns fold back into the workflow contract that governs all future phases. The tool gets measurably better the more you use it.

Table of contents

• How it works (the herd loop)
• Watch it work — the live dashboard
• Install & quick start
• The plugin (Claude Code & Antigravity)
  • Skills
  • MCP tools
• The workflow in depth
  • Project docs
  • Milestone & phase layout
• The CLI
  • Command reference
  • Example output
• Built for small / local models
• The improvement loop
• Configuration — rexymcp.toml in full
• Under the hood — the executor
• Repository layout
• Development

How it works (the herd loop)