Summary

It asks before it rewrites.

The clarification contract is the interesting part: instead of guessing your audience and tone, the server hands back structured questions with IDs, your client asks you, and the answers come back mapped. That makes the refined prompt reproducible rather than a one-off improvisation — and the same flow degrades gracefully in clients that cannot ask questions.

What it is

Promptheus refines prompts for LLMs, and ships an MCP server so the refinement happens inside your client. It detects whether a task needs clarification or direct optimisation, asks targeted questions when it needs them, and returns a prompt you can run.

What you get

`refine_prompt` — refinement with an optional clarification round; it either returns the refined prompt or a set of questions plus an `answer_mapping`, which you answer through your client's own question tool and send back
`tweak_prompt` — a surgical edit to an existing prompt from a description such as "make it shorter"
`list_models` and `list_providers` — which providers are configured and which models they expose
`validate_environment` — checks configuration and, optionally, real API connectivity per provider
Provider and model can be overridden per call, so one prompt can be refined by whichever model you prefer

Requirements

Python 3.10 or higher. `pip install promptheus`, then start the server with `promptheus mcp` (or `python -m promptheus.mcp_server`). At least one provider key is needed — `GOOGLE_API_KEY`, `ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, `GROQ_API_KEY`, `QWEN_API_KEY` or `GLM_API_KEY` — with `PROMPTHEUS_PROVIDER` and `PROMPTHEUS_MODEL` available to pin the choice. The `mcp` package must be installed. Package version 0.3.2; the registry entry publishes 0.3.1 over stdio.

Setup effort

One command plus a key — pip install promptheus, then supply credentials