Labsco
MCP SERVER

Yellhorn MCP

by msnidal ↗

Spends a large reasoning model on a written implementation plan, files it as a GitHub issue, then judges the resulting diff against it.

Reasoning Scaffolds & Agent Workflow Engines✓ Verified
Summary
It puts an expensive model on the plan and the review, and leaves the cheap model to do the typing.

Your whole codebase goes into the context window by default — that is the design, and it is also why curate_context exists: run it first so a task only carries the directories it actually touches. Everything lands in GitHub, the plan as an issue and the judgement as a linked sub-issue, so the reasoning survives outside the chat window.

What it is

A planning and review server for coding agents. It sends your codebase to a big model to produce a detailed workplan, posts that as a GitHub issue, and later compares two git refs against the same plan.

What you get
  • create_workplan opens the GitHub issue immediately, then fills it in from a background run of the model over your codebase
  • revise_workplan re-runs that against your revision instructions and updates the same issue in place
  • get_workplan reads a plan back by issue number, and workplans are also exposed as MCP resources your client can list
  • judge_workplan compares two git refs, or a pull request, against the original plan and posts the verdict as a linked sub-issue
  • curate_context analyses the repository for a stated task and writes a .yellhorncontext file naming the directories worth sending to the model
  • Context control through .yellhornignore and .yellhorncontext, automatic chunking when a codebase outgrows the model's window, and Google Search grounding with Markdown citations on Gemini models
  • Cost estimation and usage tracking on every API call
Requirements

The GitHub CLI (gh), installed and authenticated — every workplan and every judgement is a GitHub issue. A key for the model family you pick: GEMINI_API_KEY, OPENAI_API_KEY or XAI_API_KEY. YELLHORN_MCP_MODEL selects the model (default gemini-2.5-pro, with GPT-5, o3, Grok-4 and deep-research options), YELLHORN_MCP_REASONING_EFFORT sets low, medium or high on GPT-5 models, and YELLHORN_MCP_SEARCH turns search grounding on or off. REPO_PATH points at the repository and defaults to the current directory. Install with uv pip install yellhorn-mcp and run it over stdio as uv run yellhorn-mcp.

Setup effort

One command plus a key — uv pip install yellhorn-mcp, then supply credentials