An MCP phone number is a real phone number that an AI agent controls through Model Context Protocol tools. With AgentLine, you add the MCP server at https://api.agentline.cloud/mcp with your API key as a bearer header. The agent can then provision a US or Canada number, place and answer calls, send and read SMS, and fetch transcripts as ordinary tool calls. Plans start at $20/month with a number included.
Connect the MCP server
Create an API key in the AgentLine dashboard and set it privately as AGENTLINE_API_KEY. Any MCP client that can reach a remote server works. For a client that launches local commands, such as Claude Desktop or Cursor, add this entry to its MCP configuration:
{
"mcpServers": {
"agentline": {
"command": "npx",
"args": ["-y", "mcp-remote@latest", "https://api.agentline.cloud/mcp", "--header", "Authorization: Bearer $AGENTLINE_API_KEY"]
}
}
}In Claude Code the same connection is one command. The Claude Code guide, Cursor guide and Codex guide show each client's setup and first call.
claude mcp add agentline -- npx -y mcp-remote@latest https://api.agentline.cloud/mcp --header "Authorization: Bearer $AGENTLINE_API_KEY"
What the MCP tools let the agent do
The agent skill lists 35 MCP tools, and every REST endpoint is also exposed as a tool. The ones a phone workflow uses most:
- create_agent: set the name and system prompt for the hosted voice that holds the conversation.
- Number provisioning: buy a US or Canada local number for that agent, list numbers, and reassign them.
- make_outbound_call: dial an E.164 number; the hosted voice talks and the call is recorded and transcribed.
- send_sms: send a text from the agent's number; inbound texts can be read back.
- push_call_context: give the voice live facts during a call when your runtime does a lookup.
- Contacts: create_contact, list_contacts and the related tools store up to 1,000 people per agent.
Your MCP client does not voice each turn. AgentLine's hosted voice agent runs the conversation from the prompt you set, and your agent reads the transcript afterwards. To answer mid-call questions from the agent's private context, use the relay or a webhook as described in the skill.
First call in three requests
- Ask the agent to create a phone agent with a short system prompt.
- Ask it to buy a US local number for that agent.
- Ask it to call your own phone and read you the transcript when the call ends.
Each step is a tool call that shows up in your MCP client, so you can approve purchases and calls before they run. Buying a number and placing calls cost money; keep approvals on for those tools.
MCP, skill or REST
Use MCP when the agent itself should operate the phone from a chat or coding client. Use the skill file when your runtime reads skills instead of MCP servers. Use the REST API from a backend service; the API walkthrough covers the same flow with curl. All three reach the same account, numbers and transcripts.
Cost
AgentLine's published rates are monthly plans from $20 with included US or Canada numbers, call minutes and SMS; beyond the plan, $3.00/month per number, $0.12/minute for calls (per-second after a one-minute minimum; zero-second calls are free), and $0.02 per SMS in either direction. Check pricing and your account's allowances before budgeting. Runtime hosting and external integrations can add costs.
Reviewed against AgentLine API schema and agent skill. For a broader setup without MCP, see How to Give Your AI Agent a Phone Number.