The value is that it returns a level rather than a paragraph, which is what makes it usable inside a loop: LOW passes through, MEDIUM gets a caveat, HIGH blocks the merge from a sub-agent until someone looks. Try it on a plausible-sounding API that does not exist — that is the failure mode it is built for, and the one review by eye is worst at catching.
A screening service for model output. You hand it the model name, the prompt and the response, and it returns a fabrication risk score derived from behavioural fingerprinting rather than a fact-check lookup. It covers Claude, GPT, Gemini, Grok, Llama, Mistral and others.
- `arkheia_verify` scores a response for fabrication risk
- `arkheia_audit_log` reviews your detection history
- `run_grok`, `run_gemini` and `run_together` call the model and screen the answer in the same step
- `run_ollama` does the same against a local model
- A three-level result — LOW, MEDIUM, HIGH — that an agent policy can branch on directly
- A ready-made verification protocol for `CLAUDE.md` in the README, including how to handle sub-agent output
Node.js 18+, Python 3.10+ and git. `npx @arkheia/mcp-server` sets up the Python environment and configuration; the client then runs `python -m mcp_server.server` from `~/.arkheia/mcp` with `ARKHEIA_API_KEY` set. Keys are free and provisioned by one HTTP call against your email address, and shown once. The free tier covers 1,500 detections a month with 5 concurrent, no card. Screening means the prompt and the response leave your machine. Version 1.3.0.
One command plus a key — npx @arkheia/mcp-server, then supply credentials