minimaxmcp
shaunpatterson/minimaxmcp
Quick MCP server to wire up Minimax M3 as a judge into claude code
Instalação
npx convez-ai shaunpatterson/minimaxmcp1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
minimax-mcp
An MCP server that exposes MiniMax M3 to Claude Code (or any MCP client) — primarily for use as an LLM judge: a second, independent model that reviews Claude's work.
Tools
| Tool | Purpose |
|---|---|
minimax_judge | Evaluate work (code, diffs, plans, answers) against a task and criteria. Returns a PASS/FAIL verdict, a score out of 10, and a numbered list of concrete issues with severities. |
minimax_generate | Simple one-shot prompt → response, with optional system instructions. |
minimax_messages | Multi-turn structured conversation (system/user/assistant messages). |
All tools accept optional model (default MiniMax-M3), temperature, top_p, max_completion_tokens, and include_reasoning (returns M3's thinking trace alongside the answer).
Setup
1. Get a MiniMax API key
Create one at platform.minimax.io (international platform — this server points at https://api.minimax.io/v1).
2. Build
git clone git@github.com:shaunpatterson/minimaxmcp.git
cd minimaxmcp
npm install
npm run build
3. Wire it into Claude Code
claude mcp add minimax \
-e MINIMAX_API_KEY=your-key-here \
-- node /path/to/minimaxmcp/build/index.js
Use claude mcp add --scope user ... to make it available in every project, or --scope project to share it via a checked-in .mcp.json.
Alternatively, add it to .mcp.json by hand:
{
"mcpServers": {
"minimax": {
"command": "node",
"args": ["/path/to/minimaxmcp/build/index.js"],
"env": {
"MINIMAX_API_KEY": "your-key-here"
}
}
}
}
Verify with claude mcp list — the server should show as connected. Inside a session, /mcp shows the tools.
4. Use it as a judge
Just ask in a Claude Code session:
Use minimax_judge to review the diff you just wrote against the original requirements.
Or bake it into your workflow with a skill / CLAUDE.md instruction, e.g.:
After completing any non-trivial implementation, call minimax_judge with
the original task and the final diff. Address any critical or major
issues it finds before declaring the work done.
Configuration
Environment variables:
| Variable | Required | Default | Purpose |
|---|---|---|---|
MINIMAX_API_KEY | yes | — | MiniMax API key |
MINIMAX_MODEL | no | MiniMax-M3 | Default model for all tools |
MINIMAX_BASE_URL | no | https://api.minimax.io/v1 | Override for proxies / regional endpoints |
Notes
- M3's thinking is requested with
reasoning_split: true, so<think>content never leaks into tool output; passinclude_reasoning: trueto see it. minimax_judgedefaults totemperature: 0.2for consistent verdicts; override per call if you want variance.