How to Give Your AI Agent a Phone Number (2026 Guide)

Per-minute services bill hold time like talk time. If you mostly need an agent to call businesses for you, the first row is the shortest path; the remote MCP setup for ChatGPT and Claude takes a few minutes. If your agent needs to receive calls and texts on its own number, the middle rows fit better.

When developers say they want to give their AI agent a phone number, they usually mean giving it the ability to place outbound calls, to a restaurant, a doctor’s office, an airline, or a billing department, and hold a real conversation for the user. The agent should navigate phone menus, wait on hold, talk to whoever picks up, and return a result: the network outcome, a transcript when present, and ideally a confirmation or case number.

That capability arrives one of two ways. The first is a hosted calling service exposed as a remote MCP connector, an agent skill, or a REST API. Your agent calls a tool like place_call(number, task), and the service handles telephony, voice, and the conversation. The second is building it yourself: rent a number, then connect speech-to-text, a language model, text-to-speech, and code to glue them together.

Both give your agent a working line. The difference is whether you spend minutes connecting a tool or weeks building infrastructure. The build path makes sense when you need custom voice behavior that no finished service covers.

A hosted connector ships a finished calling capability through remote MCP, with an agent skill or REST API as fallbacks. A compatible host connects through OAuth, discovers the calling tools, and can place calls from its next turn. ClawCall is one example: its OAuth-protected MCP server at api.clawcall.dev/mcp works with Claude, Codex, Cursor, Grok Bot, and other compatible hosts, with a skill and REST API for the rest. In ChatGPT you skip the endpoint: the official ClawCall plugin is listed in ChatGPT’s plugin directory, and installing it connects the same server. OpenAI dots use that same installed plugin.

A number-for-your-agent service sits in between. AgentPhone, AgentLine, and Bland’s Agent Phone Plan give the agent its own number for calls and texts, billed per minute or per month, and you decide what the agent says.

Building your own means platforms such as Vapi, Retell, or Bland, or a raw carrier such as Twilio. You bring the model, pick a voice, wire telephony, and design the dialog. Pricing is per minute and setup takes days to weeks.

If your goal is to give an existing coding or chat agent the ability to call businesses, the hosted route is faster and cheaper. If you are building a voice product with its own brand and dialog design, start with a platform.

Here is the hosted path, using ClawCall as the example:

1. In an MCP-compatible host, add https://api.clawcall.dev/mcp as a remote MCP server.
2. Sign in to ClawCall and approve the calls:read and calls:write scopes. OAuth keeps the ClawCall API key out of the AI host.
3. Let the host discover the tools. It finds twelve bounded tools, including the required get_calling_guide tool, call placement, status polling, the transcript when present and temporary recording access when available, history, balance, and hangup.
4. Run a read-only balance check to confirm the connection.
5. Place a controlled call to a number you own and read the transcript when present.

Once connected, ask normally. The host can pick ClawCall on its own when a task needs a call. Claude Code, Codex, Cursor, and Grok Build accept the same remote MCP URL in their own configuration. For hosts without remote MCP, use the packaged skill or the REST API. None of these paths needs an SDK, a Twilio account, SIP configuration, or a copied API key in chat.

For a host without remote MCP, the REST API is the lowest-friction way to add phone calls to an agent. The contract is small.

The agent sends POST /call with a JSON body containing the destination number in E.164 format and a plain-English task that includes the facts the agent will need on the line, such as the caller’s name, account numbers, and dates. The server validates the number (US +1 only), takes an outbound line, starts the dial, and returns a call_id immediately. The API is fire-and-poll, not blocking.

Your agent then calls GET /call/:id every few seconds and watches lifecycle move from queued to dialing to answered to finalized. Once lifecycle is finalized, the response includes outcome plus nullable talk_seconds, transcript, and recording_url fields. Read the transcript only when it is present, and copy recording audio only when recording_url is present.

A polling loop:

1. POST /call and keep the call_id.
2. Wait three seconds, then GET /call/:id.
3. Repeat until lifecycle is finalized.
4. Read outcome, and the transcript when present.

Calls with long holds take as long as the hold, so keep the interval modest. The full schema lives in the API tutorial and the developer reference, licensed CC BY 4.0.

Real calls are messy. The number routes through a phone menu, waits on hold, transfers to another department, and then asks for a code the agent does not have. A good calling service handles the first three by default.

• Phone menus work because the agent hears the options and presses the right key with the send_dtmf tool.
• Hold time is absorbed silently. The agent stays on the line and resumes when a person returns.
• Transfers are followed, because the call stays connected.
• Verification codes need a person. The fix is loop-in: with loop_in_user enabled on the call, the agent can call your verified phone and connect you so you can read the code or speak to the rep.

Every plan allows one outbound call at a time per account. Trial and Unlimited calls use shared lines. A second call while one is running returns concurrent_call_limit. Wait for the current call to finalize before placing the next. If all shared lines are busy with other accounts, number_pool_exhausted means you can retry shortly. Unlimited Reserve and Reserve Plus calls always use your reserved number, one call at a time. On shared-pool calls, looping you in uses a second shared line while connected. Loop-in is also the right pattern for any moment that needs a real person: approving a price, agreeing to terms, or handling an upset rep.

An illustrative example: an agent calling a utility to dispute a bill handles the menu and the hold music, then loops the user in only when the rep asks for the last four digits of their SSN. A 40-minute call becomes a 40-second human moment.

Pricing comes in two shapes. Per-minute services charge for connected time: AgentLine lists $0.10 a minute, AgentPhone $0.13 to $0.22, and Retell $0.07 to $0.31. Flat-rate services charge a fixed monthly fee.

ClawCall is flat:

• Free trial: 30 calls and 30 minutes, whichever lasts later, with no credit card.
• Unlimited, $4.99 a month: unlimited calls from the shared outbound number pool.
• Unlimited Reserve, $8.99 a month: adds one private reserved caller-ID number.
• Unlimited Reserve Plus, $14.99 a month: adds an AI inbound assistant on that reserved number.

There is no per-minute meter on any plan.

Limits worth knowing: ClawCall dials US +1 numbers only, in English only. The agent always discloses that it is an AI when asked, leaves voicemail only when instructed, and does not send bulk or robocall-style promotions, while individual outreach calls are allowed. There is no public outbound SMS endpoint and no HIPAA, PCI, or SOC 2 attestation. ClawCall deletes transcripts 7 days after a call ends and recordings 10 minutes after a call ends, and a call can finish without a transcript. If your agent needs to receive texts, or call outside the US, one of the number services in the table fits better.

If you want your AI agent to have a working phone today:

1. In ChatGPT, install the official ClawCall plugin from ChatGPT’s plugin directory: select Plugins, search “ClawCall”, select Install plugin, then Allow connection. In Claude, Codex, Cursor, Grok Bot, or another compatible host, add the hosted MCP connector. If the host does not support remote MCP, install the packaged skill or use the REST API.
2. Place one test call to your own mobile number. Listen to the voice, confirm the AI disclosure, and read the transcript when present.
3. Wire the calling tool into the workflow you care about: scheduling, follow-ups, hold time, or bill disputes.
4. Use the free trial of 30 calls and 30 minutes, whichever lasts later, then decide whether the $4.99 Unlimited plan covers your usage.

The agent setup guides cover authentication in more depth, and the Hold For Me workflow shows loop-in end to end. For the build-your-own platforms, see ClawCall vs Vapi, ClawCall vs Bland, and ClawCall vs Retell.

• Bland AI Agent Phone Plan
• AgentPhone docs: Usage and billing
• AgentLine pricing
• Twilio Voice pricing for the United States
• Retell AI pricing