MCP server

Let coding agents lint prompts and check LLM calls in the code they write.

pip install "counterfix[mcp]"
claude mcp add counterfix -- counterfix mcp        # Claude Code; any MCP client works

Tool	Does	Model calls
lint_prompt	rules L1-L7 on a prompt	none (judge rules if judge=true)
diagnose_call	why an answer failed	none (judge checks if judge=true)
fix_call	a fix that beat a plain retry on re-run	yes, with your keys
compare_profiles	failure rates of two models	none
route_models	cheapest model per task type	none

The server reads a .env in the folder it starts in, the same way the CLI does. It runs on stdio.