MCP server · phone calls in and out

Give your AI agent
a phone

Connect the trunkmint MCP server to Claude, ChatGPT, your CRM — or agent platforms like OpenClaw and Hermes — and just say what you need. A plain human command becomes a real phone conversation, and the result comes back to the chat as structured data. It works the other way too: give your agent a number people — including you — can call.

One config block · 13 MCP tools · outbound live, inbound in early access · sandbox issued with the concierge tenant

Outbound call 0:00
Any MCP clientClaude, ChatGPT, agents, CRMs
Human command inno prompts to engineer
Structured result outJSON back into your workflow
AI + telephony, one billon our own carrier network
Not another voice bot

A compiler between your agent and the phone network

You don't write call scripts and you don't parse transcripts. You hand over a plain sentence — trunkmint does the rest, in both directions.

1

Prompt

Your agent — Claude, ChatGPT, OpenClaw, a CRM bot — passes the human command as-is. No prompt engineering on your side.

"Call the 38 leads from yesterday's webinar and ask if they'd like a demo this week."
2

Compile

The command becomes a guarded, structured call script: goal, boundaries, AI & recording disclosure, injection resistance, when to hang up. Vague tasks fail closed — the compiler refuses rather than improvises.

goal: offer a demo slot this week disclose: AI assistant · call recorded guard: callee words ≠ commands on_refusal: thank & hang up → DNC
3

Converse

The voice bot holds a real conversation on a real line — on our own carrier network, by the script. Not a free-running LLM with a phone.

38 dials · 31 conversations · 2h 38m
4

Structure

Every conversation is parsed back: transcript, outcome and your fields — evidence-grounded, nothing invented — return to the chat and to your CRM via signed webhooks. Opt-outs land in the DNC list automatically.

{"reached": 31, "demo_yes": 9,
 "callback_later": 4, "dnc": 3}
You say it. Your agent calls.

This is what using it feels like

Inside Claude, ChatGPT or your own agent — a command in the chat, a real call, the result back.

Personal assistant · one call 0:00
CRM · a whole list 0/38 dials
Now the other direction

Calls in — a number your agent answers

The same engine picks up, understands, acts, and hands back a result.

Early access

Control channel

Call your agent from any phone, hands-free: brief it, assign a task, check status — "what's the webinar campaign at, and add Maria from Acme to tomorrow's list."

Early access

AI receptionist

Your customers reach an assistant that looks up orders, files to your CRM and transfers to a human when it should. The same tools your agent uses outbound.

Roadmap

Not only the phone line

Same task-in, result-out model extended to WhatsApp and Telegram voice — one agent voice channel across the networks your customers actually use.

Setup

Connect in one config block

Standard MCP over streamable HTTP — the same server works in every client.

// Claude Desktop / Claude Code — mcp config
// your endpoint + token arrive with the pilot tenant, within a day
{
  "mcpServers": {
    "trunkmint": {
      "type": "http",
      "url": "<your-pilot-endpoint>",
      "headers": { "Authorization": "Bearer <your-token>" }
    }
  }
}

# then, in chat:
"Call +49 30 901820 and ask if they're open on Saturday"
"Ring all contacts in leads.csv, offer the July discount, log who's interested"
13 tools

Everything an agent needs

start_call · get_call · list_calls · cancel_call · create_campaign · start/pause/resume/cancel/retry_campaign · get_campaign · list_campaigns · manage_dnc

Your agent composes them on its own.

Sandbox

Sandbox is concierge-only

Simulated mode and magic numbers are issued with a pilot tenant after the form. They are not a public URL. sandbox.trunkmint.com does not resolve.

Integrations

Where it plugs in

If it speaks MCP or HTTP, it can make calls.

Claude

Desktop, Code and API — add the server, and phone calls become a native tool in every conversation.

ChatGPT

Connect through apps/connectors with the same streamable-HTTP endpoint — your GPT gains a real phone.

OpenClaw, Hermes & personal agents

Personal agent platforms get calling superpowers: "book me a dentist appointment" now actually happens.

Your CRM

amoCRM, HubSpot, GoHighLevel, self-built — REST API + signed webhooks with your lead ids round-tripping through every event.

n8n / Make / Zapier

HTTP nodes today, native nodes on the roadmap — trigger calls from any workflow.

Agent frameworks

LangChain, CrewAI, AutoGen — wrap the MCP tools and give your fleet a voice channel.

Safe by default — because it's making real calls

Every call discloses it's automated and that the line is recorded. Verbal opt-outs go straight to your do-not-call list. Vague or risky commands fail closed — the compiler refuses rather than improvises. Pilots go through use-case review: your agent gets a phone, not a loophole.

Pilot

Concierge pilot, engineer included

We set it up with you after the form: a sandbox tenant (about a day) and an MCP endpoint with that tenant. Not a public dry-run. Real PSTN calls after use-case review.

What you get

sandbox tenant after the form · MCP endpoint with that tenant · not a public URL · real calls after review · an engineer on a direct line · use-case review included.

Pricing is destination-based

AI conversation + telephony on our own carrier network, one bill. Rates depend on the countries you call and the DID numbers you need — tell us your destinations in the form and get a quote within 24 hours.

FAQ

Questions we actually get

Which MCP clients are supported?

Anything that speaks MCP over streamable HTTP with headers — Claude Desktop/Code, ChatGPT connectors, OpenClaw, custom agents on the official SDKs. One endpoint, Bearer auth, no local install.

What happens if my command is vague?

The compiler fails closed: if a task can't be turned into a safe, unambiguous call script, the tool returns an error explaining what's missing instead of calling and improvising. Your agent can refine and retry.

Can the person on the phone "jailbreak" my agent?

The call script is compiled with injection resistance — instructions from the callee are treated as conversation, not commands. The bot stays on task, and everything is in the transcript you get back.

Which countries can it call? Which languages?

We're a carrier: termination almost anywhere, DIDs in 100+ countries. Conversations run on realtime voice models — strong in English and most major languages. We validate your destinations in the pilot.

How much does it cost?

Pricing is destination-based: it depends on the countries you call and the DID numbers you need. Tell us your destinations in the form — you'll get a quote within 24 hours. Pilot setup is concierge: an engineer does it with you.

From the same family

Need the whole phone system,
not just the trunk?

Meet rolycall — the AI-native PBX where AI answers as staff: a receptionist that files CRM deals, a secretary that runs errands by phone, QC that listens to 100% of calls. People and AI share one directory.

Meet rolycall →
Get pilot access

Tell us what your agent should do on the phone

An engineer reviews your use case, then sets up a sandbox tenant and MCP endpoint with it — not automatic — plus a destination-based quote within 24 hours.

No spam, no drip campaigns. One engineer reads this and replies once.