Safety & anti-spam

How DialMCP keeps AI phone calls honest — and rare.

Giving AI agents the ability to place phone calls is only defensible if abuse is structurally hard. DialMCP's protections are enforced server-side on every call — they are not model instructions an agent could ignore, and not settings a user can switch off.

Every call announces itself

• Disclosure first. The first thing the agent says to any human is that it is an AI assistant, who it is calling on behalf of, and that the call is recorded. This happens before the ask — even if the callee starts the conversation. Compliance is independently verified against the transcript after every call and logged.
• It never denies being an AI. Asked directly, it confirms. If the callee objects to talking to an AI or to being recorded, the agent apologizes and ends the call immediately — it does not argue or persuade.
• Real caller ID. Calls present the user's own SMS-verified phone number. Callees can call back a real person. There are no spoofed, rotating, or anonymous numbers.
• Every call is recorded and transcribed, and the recording notice is part of the disclosure. There is a complete audit trail: who called whom, when, what was said, and how it ended.

Hard rate limits

Volume limits make bulk calling impossible by construction:

• 1 call at a time per user — no parallel dialing, ever.
• 3 calls per hour and 10 calls per day per user.
• 2 calls per day to the same destination — an agent cannot pester a number.
• 10-minute hard cap per call (6-minute default), enforced by a server-side watchdog.
• Burst limiting on top, so even the hourly allowance can't be fired in one second.

When and where calls are allowed

• Calling hours: calls may only start between 8:00 and 21:00 in the destination's local time zone, aligned with US telemarketing law (TCPA). Unknown time zones are treated conservatively.
• US and Canada only. No international destinations.
• Non-bypassable blocklist: emergency services (911 and related), premium-rate lines (900/976), and high-fraud number ranges can never be dialed, under any configuration.
• Line-type screening: every destination is looked up before dialing, and the line type is factored into whether — and on what legal basis — the call proceeds. Each call's consent rationale is written to a compliance log.

Do-not-call, honored permanently

• If a callee says any version of “don't call me again”, the agent confirms it on the spot and the number is added to a permanent suppression list. A second, independent check re-reads every transcript after the call and catches opt-outs the agent may have missed.
• The suppression list is checked before every call, for every user — one opt-out blocks the number platform-wide.
• Anyone can opt out a number directly — no account needed — at mcp.dialmcp.com/opt-out.

Objective screening

• Before dialing, the stated objective of every call is screened by an independent AI reviewer for harassment, deception, fraud, or other abuse. Objectives that fail the screen are rejected with a reason.
• The voice agent only states facts from the task the user provided — it is instructed never to invent prices, names, or details, and to say “I don't know” and offer a follow-up instead.
• Voicemail gets nothing: if a machine answers with a mailbox greeting, the agent hangs up without leaving a message.

The kill switch

Operationally, the platform has a global kill switch that stops all outbound calling immediately, and per-number controls for finer-grained intervention. If something looks wrong, calling stops first and questions get asked second.