We build UIRIX, a paid product that lets Claude or ChatGPT place real phone calls through a remote MCP server. A chat model with a phone can do harm quickly, so most of the design is about what the server refuses. These are the five rules, and how each one looks to the model.
1. No call without a recorded yes
The call tool takes a confirmed field. Anything other than true is refused, and nothing is dialed or billed. The model has to repeat the number, the task and the timing to the user and get a yes first. A sentence in the system prompt was not enough: the server has to refuse.
2. A task that is missing something essential is not dialed
"Book a table" with no party size or date comes back as "needs input", with the list of what is missing. The question goes to the user before the call, and the business is not kept on the line while the model guesses.
3. The agent says what it is
It calls from its own phone number and opens as the user's AI assistant, calling on their behalf. It never speaks as the user.
4. Phone menus are listened to, then pressed
The agent waits for the menu to finish and presses only the option that leads to the task. Hold has a budget: 3 minutes by default, 10 at most, because hold minutes cost the user money. When the menu or the queue wins, the result says so.
5. Placing calls is its own permission
It is off when the user connects the server, and it is switched on separately from reading or managing anything else.
What comes back
The model gets a report built from the transcript: what was agreed, with whom, and what is still open. One yes can cover up to three businesses for the same errand, for price quotes, and never more.
What it does not do
It does not pay for anything, find a business on its own, or call for sales, collections or emergencies.
If you want to read the tool descriptions
claude mcp add --transport http uirix https://dashboard.uirix.com/api/ext/v1/mcp
Sign-in is OAuth. Placing calls needs a UIRIX phone number, which needs a saved card, and the plan is $9.99 a month. Details: https://www.uirix.com/en/integrations/claude-chatgpt
If you run a tool that acts in the real world: where does your server refuse, and where do you trust the model?
Top comments (0)