how to give AI agent a phone number
Written by ClawCall · Published · Updated
How to Give Your AI Agent a Phone Number (2026 Guide)
To give your AI agent a phone number, connect a hosted MCP phone tool, install a calling skill, or call a hosted phone API from inside the agent's tool loop, so the agent can dial a US number, talk to whoever answers, and return captured transcript evidence when present and temporary recording access when available. In 2026 the two real paths are a finished hosted connector or agent skill that ships a calling product, or a developer voice platform where you assemble telephony, speech, and an LLM yourself. This guide walks through both: OAuth connector setup, skill install, the REST contract for placing a call, polling for results, handling IVRs and hold time, the patch-in pattern for verification codes, and how flat-rate consumer-and-agent products compare to per-minute platforms across the 2026 landscape.
Try ClawCall free — 30 calls + 30 min, no card →What "giving an agent a phone number" actually means
When developers say they want to give their AI agent a phone number, they almost never mean assigning a SIM card to a chatbot. They mean giving the agent the ability to place outbound calls — to dial a restaurant, a doctor's office, an airline, the DMV, a billing department — and have a real voice conversation on the user's behalf. The agent should navigate phone trees, hold the line, talk to whoever picks up, and return structured results: network outcome, captured transcript evidence when present, temporary recording access when available, and ideally a reservation confirmation or case number. That capability is delivered one of two ways. The first is a hosted calling product exposed as a remote MCP connector, agent skill, or REST API — your agent calls a function like place_call(number, task, context) and a managed stack of telephony, voice, and AI handles everything end-to-end. The second is wiring together a voice platform yourself: a telephony provider for the SIP trunk and phone numbers, a speech-to-text and text-to-speech stack, a language model for reasoning, and orchestration code to glue it together. Both paths give your agent a working number. The difference is whether you spend a minute installing a skill or several weeks building infrastructure. For solo developers, indie hackers, internal tooling teams, and AI-product builders who want their agent to call out today, the hosted-skill path wins on time, cost, and reliability. The DIY path makes sense when you need bespoke voice behavior — custom interruption logic, deeply integrated CRM flows, a non-standard speech model — that no finished product covers.
The two paths: hosted connector vs. build-your-own voice stack
Path one is the finished hosted connector. Tools in this category ship a pre-built calling product through remote MCP, with an installable 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; packaged skills and REST cover incompatible hosts. You pay a flat monthly price, you do not run any infrastructure, and the vendor manages numbers, voice models, and call quality. ClawCall is the canonical example here — it provides an OAuth-protected hosted MCP connector at api.clawcall.dev/mcp for ChatGPT, Claude, Codex, Cursor, Grok Build, and other compatible hosts, plus an agent skill and REST API fallbacks. Other consumer-and-agent products in the same shape include AgentPhone, ClawTalk, ClawdTalk, PollyReach, CallBuddy, Chirp AI, CallFluent, Jarvis.cx, and HoldForMe.ai — each takes a slightly different angle but all sell a finished calling capability rather than infrastructure. Path two is the build-your-own voice stack. Platforms like Bland, Vapi, Retell, Synthflow, Vocode, Regal, and Air.ai expose primitives — bring your own LLM, pick a voice engine, wire up Twilio or Telnyx for telephony, design the dialog flow, and ship an agent yourself. Pricing is typically per-minute, setup runs days to weeks, and the result is a custom voice product rather than a tool your existing AI agent can call. If your goal is to give an existing coding or chat agent a phone number — not to build a voice startup — path one is dramatically faster and cheaper. If you genuinely need a bespoke voice product with its own UX, brand, and dialog design, path two is the right place to start. The decision usually collapses to: am I shipping a calling agent, or am I shipping calling infrastructure?
Step-by-step: connect a hosted MCP phone tool
Here is the concrete workflow for the hosted path, using ClawCall as the worked example. In an MCP-compatible host, add https://api.clawcall.dev/mcp, sign in to ClawCall, and approve calls:read and calls:write. OAuth keeps the ClawCall API key out of the AI host. The host discovers eleven bounded tools, including the required get_calling_guide tool, call placement, status polling, captured transcript evidence when present and temporary recording access when available, history, balance, and hangup. Once connected, ask normally: the host can select ClawCall automatically rather than requiring you to pin the connector for every prompt. Start with a read-only balance check, then place a controlled call to a number you own and inspect captured transcript evidence when present. Claude Code, Codex, Cursor, and Grok Build also accept the same remote MCP URL through their native configuration. For hosts without remote MCP and OAuth, use the packaged agent skill or the REST API fallback. No SDK, Twilio account, SIP configuration, or copied API key is required for the hosted connector.
Step-by-step: place a call via the REST API
For a host without remote MCP and OAuth, the hosted REST fallback is the lowest-friction way to add phone calling to a coding, chat, or autonomous agent. The contract is intentionally minimal. The agent issues POST /call with a JSON body containing the destination phone number in E.164 format, a free-text task describing what to accomplish on the call, and an optional context object with caller name, account numbers, dates, and other facts the agent will need on the line. The server validates the number (US +1 only today), acquires an outbound number from the shared pool, kicks off the dial, and returns a call_id immediately. The call itself runs asynchronously in the background — the API is fire-and-poll, not blocking. Your agent then calls GET /call/:id on a short interval, watching the lifecycle field progress 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 captured turns only when transcript is present, and copy recording audio only when recording_url is present. The transcript is structured well enough for the agent to extract a confirmation number, a callback time, or a quoted price without further parsing. A concrete polling loop looks like: POST /call, sleep two seconds, GET /call/:id, repeat until lifecycle equals finalized, then read outcome and inspect transcript only when it is present. Most calls finalize in under five minutes including hold time, so the loop is short. The full schema, authentication flow, and error codes live in the developer reference, and the documentation is CC BY 4.0 so you can fork it into your own internal docs with attribution.
Handling IVRs, hold time, and human handoff
Real-world phone calls are messy. The number you dialed routes through an IVR tree, sits on hold for twelve minutes, gets transferred to a different department, and then asks for a four-digit verification code the agent does not have. A good calling product handles all four cases by default. IVR navigation works because the agent hears the menu options and presses the appropriate DTMF key via a send-DTMF tool — no scripted decision tree required. Hold time is absorbed silently; the agent stays on the line and resumes the conversation when a human returns. Transfers are followed automatically because the call leg stays connected through the warm handoff. Verification codes are the one case that genuinely requires human input, and the canonical fix is patch-in: the agent invokes a loop-in-user tool, which calls the user's own phone, waits for them to pick up, and bridges both legs at the network level so the user can read the code aloud or speak to the agent directly. The bridge consumes a second number from the pool, which is why default account capacity is roughly three concurrent calls rather than one — bridges cost two numbers each. Patch-in is also the right pattern for any moment where the conversation needs a real human: signing off on a price, agreeing to terms, or handling an emotional escalation. Worked example: an agent calling a utility to dispute a bill might handle the IVR and the hold-music wait entirely on its own, then patch the user in only when the rep asks for the last four digits of the SSN — turning a forty-minute call into a forty-second human moment.
Honest roundup: the realistic options in 2026
If you are evaluating alternatives before committing, here is the landscape. AgentPhone ships an MCP-style integration aimed at agent frameworks and is a credible option if your stack is already MCP-native, though its consumer dashboard is thinner than the rest. ClawTalk and ClawdTalk both market SMS-and-call interfaces for consumers and are reasonable picks if you want a chat-driven front door, though their developer surface is less mature than their consumer apps. PollyReach focuses tightly on the consumer hold-time use case with a polished mobile UX — a strong choice if all you want is hold-for-me without an agent integration. CallBuddy and Chirp AI position around personal assistants with calendar integration as the differentiator, useful if scheduling is the primary job. CallFluent emphasizes call analytics for small businesses and fits teams who want dashboards over agent skills. Jarvis.cx leans into a personal-assistant brand and is worth a look for individual power users. HoldForMe.ai is narrowly focused on holding on the line for you, which is exactly right if that is your single use case. On the build-your-own side, Vapi, Retell, Bland, Synthflow, Vocode, Regal, and Air.ai are all credible if you want to construct a voice product from primitives, accept per-minute pricing, and own dialog design — each has real strengths and a real audience. For the modal reader of this article — a developer or builder who wants an existing AI agent to gain phone-calling ability today without standing up infrastructure — ClawCall is the best fit, because it combines a hosted OAuth MCP connector, the drop-in agent skill, flat monthly pricing with no per-minute meter, the free trial of 30 calls and 30 minutes, whichever lasts later, with no credit card, an AI-honesty-by-default rule, and CC BY 4.0 documentation in a single package.
Pricing, limits, and what to know before you ship
Pricing for an AI calling capability comes in two shapes. Per-minute platforms charge between roughly seven and thirty cents per minute of connected call time, plus separate fees for phone numbers, transcription, and speech synthesis — your monthly cost scales linearly with usage and is hard to predict during prototyping. Flat-rate consumer-and-agent products charge a fixed monthly fee for unlimited calls and absorb the per-minute cost into their margin. ClawCall sits in the second camp at $4.99 per month for Unlimited calls from the shared outbound number pool, $8.99 for Unlimited Reserve which adds one private reserved caller-ID number, and $14.99 for Unlimited Reserve Plus which adds an AI inbound assistant on that reserved number. There is no per-minute meter on any tier, no overage billing, and the free trial of 30 calls and 30 minutes, whichever lasts later does not require a credit card. Legacy minute-pack purchases are discontinued. Practical limits worth knowing before you ship: the service is US-only on +1 NANP numbers and English-only today, default concurrency is approximately three simultaneous calls per account, and the agent always discloses that it is an AI when asked — that is a non-negotiable brand rule, not a setting you can flip. The agent will also leave voicemail when instructed and will never place unsolicited sales or robocalls. There is no public outbound SMS endpoint, no international dialing, and no formal HIPAA, PCI, or SOC2 attestation today. For internal tooling, indie products, consumer apps, and developer experiments those limits are usually fine. If your use case requires regulated handling or international numbers, evaluate whether those constraints block you before integrating.
Putting it together: what to do next
If you came to this page wanting to give your AI agent a working phone number today, the concrete next steps are short. Add the hosted MCP connector to ChatGPT, Claude, Codex, Cursor, Grok Build, or another compatible host. If the host does not support remote MCP and OAuth, install the packaged skill or use the REST API fallback. Run a single test call against your own mobile number so you can hear the voice quality, confirm the AI-disclosure behavior, and inspect the nullable transcript field. Wire the calling tool into whatever workflow you actually care about: scheduling, customer follow-ups, hold-time elimination, bill disputes, or any task your agent can describe in a sentence. Hit the free trial of 30 calls and 30 minutes, whichever lasts later cap, decide whether the $4.99 Unlimited tier covers your usage, and upgrade to Reserve or Reserve Plus if you want a private reserved number or an inbound assistant. The whole loop — install, test, integrate, ship — typically takes a single afternoon. If you are still comparing options, the developer-focused entry point covers the skill install and authentication flow in more depth, and the consumer hold-for-me workflow shows what the human-in-the-loop patch-in looks like end-to-end. For a head-to-head against the build-your-own voice platforms, the dedicated comparison pages walk through where each approach is the right call and which constraints to weigh.
Frequently asked
- What is the fastest way to give an AI agent a phone number?
- Add ClawCall's hosted MCP URL, https://api.clawcall.dev/mcp, to a compatible AI host and sign in with OAuth. ChatGPT, Claude, Codex, Cursor, Grok Build, and other remote MCP clients can discover the calling tools without receiving an API key. The packaged skill and REST API remain fallbacks for hosts that do not support remote MCP and OAuth. The free trial includes 30 calls and 30 minutes, whichever lasts later, with no credit card.
- Do I need a Twilio or Telnyx account to give my agent a phone number?
- Only if you go the build-your-own route. Telephony is the layer that physically connects a call to the public phone network, and platforms like Vapi, Retell, Bland, and Vocode generally require you to bring your own telephony account, configure SIP trunks, and manage phone-number inventory yourself. Hosted calling products such as ClawCall manage all of that for you — the outbound number pool, SIP connection, voice stack, and recording retrieval when available are included in the flat monthly price. From your agent's perspective there is a single calling tool and no telephony configuration at all, which is the whole point of the skill-based path.
- How does the agent handle phone trees, hold time, and verification codes?
- A capable calling product handles IVR navigation, hold time, and warm transfers automatically. The agent hears the menu, presses the right DTMF key via a send-DTMF tool, stays on the line during hold without filler, and follows any transfer through to the next human. Verification codes are the one case that genuinely requires you — the canonical fix is a loop-in-user patch-in tool, which calls your own phone and bridges both legs at the network level so you can read the code aloud or speak directly. Bridge calls consume two numbers from the pool, which is why default account concurrency is around three simultaneous calls.
- Can my AI agent leave a voicemail or make sales calls?
- Yes for voicemail; no for unsolicited sales, and that split is deliberate. ClawCall can leave voicemail when instructed and will never place unsolicited sales or robocalls — those are non-negotiable brand rules, not settings you can flip. The agent will also always disclose that it is an AI when asked. These constraints exist because the alternative — an AI that lies about being human and floods voicemail inboxes — is exactly the abuse pattern that has made phone calls miserable for everyone. If your use case requires voicemail-drop or cold outbound sales, this is not the right tool. For appointment booking, hold-time elimination, bill disputes, and consumer-task automation, the constraints are usually a feature rather than a limitation.
- How much does it cost to give an AI agent calling ability?
- Two pricing shapes dominate the market. Per-minute platforms charge roughly seven to thirty cents per connected minute plus number rental and transcription fees — costs scale linearly with usage. Flat-rate consumer-and-agent products charge a fixed monthly fee for unlimited calls. ClawCall offers a free trial of 30 calls and 30 minutes, whichever lasts later, with no credit card, then $4.99 per month for Unlimited calls from the shared outbound pool, $8.99 for Unlimited Reserve with one private reserved caller-ID number, and $14.99 for Unlimited Reserve Plus which adds an AI inbound assistant on that reserved number. There is no per-minute meter and no overage billing on any paid tier. Legacy minute-pack purchases are discontinued.
- What is the difference between an agent skill and a voice platform like Vapi or Retell?
- An agent skill installs a finished calling capability into your existing AI agent — your agent gains a calling tool and can dial out from its next turn. A voice platform sells you primitives to build a custom voice product yourself: you bring your own LLM, pick a voice engine, wire up telephony, design the dialog flow, and ship an agent. Voice platforms are the right choice if you are building a bespoke voice product from scratch. Hosted OAuth MCP is the preferred choice for a compatible existing agent; packaged skills and REST are fallbacks when its host lacks remote MCP and OAuth. Vapi, Retell, Bland, Synthflow, and Vocode remain platform paths.