Everything travels inside the one call — paths, query, model and mode — because nothing is remembered between calls, which is also what makes it cheap to send the same question to a different model when an answer looks wrong. Two of its own conventions are worth adopting: pass output_file for anything report-shaped, and on a hard question spawn three parallel calls with the query worded differently each time.
A single-tool server. One consultation call takes absolute file paths, a query, a model ID and a reasoning mode, assembles the files into one context, and returns that model's answer.
- Any OpenRouter model named per call, so the file bundle can go to a 1M-context Gemini or a 2M-context Grok rather than to whatever you are chatting with
- Three reasoning modes — fast for quick answers, mid for code reviews and bug analysis, think for security audits and complex refactoring
- File lists that mix individual paths with wildcards in the filename, such as /path/src/*.py
- output_file writes the answer to disk and returns only the path, so a long report never lands in the transcript
- Zdr routes the call only to endpoints with a Zero Data Retention policy
- Openrouter/fusion sends one query to a panel of frontier models and returns a judge's synthesis, with the mode setting the panel's research budget
- __pycache__, .env, secrets.py, .git and node_modules are skipped when files are collected
- The response footer reports what the call cost in USD
An OpenRouter API key, passed on the launch command itself: uvx consult7 your-openrouter-api-key. That needs uv on the machine; nothing else is installed by hand. Paths must be absolute, wildcards belong in the filename rather than the directory, and the size a bundle may reach follows the context window of the model you name. MIT.
One command plus a key — uvx consult7 your-openrouter-api-key, then supply credentials