An agent asked to check its own work tends to agree with itself; routing the review to Deepseek's reasoning model breaks that loop, which is the entire premise. Costs land on your Deepseek account per call, and the timeout defaults to 30000 milliseconds — worth raising before you send a long file for review.
An MCP server that gives an LLM agent access to a different model for review work. Each tool frames a specific kind of critique — code, design, prose, or feature ideas — and routes it to Deepseek's reasoning model through the Deepseek API, so the answer comes from a second model rather than the one already in the conversation.
- A code review over a file with the language named, covering bugs, style and best practices, performance and security — `code_review`
- A design critique for a named design type such as a web UI, covering patterns, accessibility and consistency — `design_critique`
- Writing feedback for a named writing type such as documentation, covering grammar, clarity and structure — `writing_feedback`
- Feature enhancement brainstorming around a concept, with feasibility and user-value assessment — `brainstorm_enhancements`
- A second-opinion path for approaches you are unsure about, with worked examples in the repository for each tool showing the request format and a sample response
A Deepseek API key. `DEEPSEEK_API_KEY` and `DEEPSEEK_MODEL` are both required, with `deepseek-reasoner` as the documented model; `DEEPSEEK_MAX_TOKENS` defaults to 8192, `DEEPSEEK_MAX_RETRIES` to 3, and `DEEPSEEK_TIMEOUT` to 30000 milliseconds. Build from a clone with `npm install` and `npm run build`, then register `build/index.js` with `node` and put the variables in the client's `env` block. The package is `mentor-mcp-server` 1.0.0, Apache 2.0 licensed.
Build from source — clone the repository and build it, then point your client at the binary