Koma Gate MCP

A local MCP server that exposes Koma Gate's LLM-based classifier so an assistant can check untrusted input for prompt injection, jailbreaks and scope before it.

What you can do

Classify untrusted text: call classify_input with the user input and an optional preset (general, code, support or reference) to get a verdict back, including whether the input is allowed, whether it is in scope, a reason, the preset used and the model that judged it.

When to use it

Useful when an agent handles untrusted user text and wants a pre-check before acting on it, for example in support, coding or reference workflows where a preset can be chosen. It is a heuristic gate, not a guarantee, so pair it with other safeguards.

Requirements

Runs locally over stdio, typically via npx with the npm package. Needs an LLM provider selected with KOMA_PROVIDER (openai, anthropic, google, deepseek or ollama; default google) and the matching API key: GEMINI_API_KEY, OPENAI_API_KEY, ANTHROPIC_API_KEY or DEEPSEEK_API_KEY. KOMA_MODEL optionally overrides the model. Network access to the chosen provider is required.