Meet Modi still remembers the first time he tried to get a phone number for an AI agent he was building. Not a shared number funneling through some messaging queue, but an actual, individual phone number—the kind you'd give to a restaurant when placing an order or to a doctor's office for callback reminders.
What followed was weeks of paperwork. Brand verification through The Campaign Registry. Decoding carrier trust scores. Waiting for approvals that never quite arrived on schedule. The kind of bureaucratic maze that makes sense when you're AT&T managing millions of human customers, perhaps less so when you're a developer trying to let a software agent make a dentist appointment.
"Legacy telecom infrastructure was built for human-to-human communication," Modi says, a reality that becomes painfully obvious the moment autonomous systems try to plug into it.
His solution? AgentPhone, a San Francisco startup that emerged from Y Combinator's Spring 2026 batch with a pitch that sounds almost too simple: one API to provision phone numbers for AI agents, complete with SMS and voice support, with iMessage, WhatsApp, and RCS mentioned as future additions. No A2P 10DLC registration gauntlet. No multi-week waiting periods. Just spin up a number the way you'd spin up a cloud server.
The company launched in May 2026, and if the timing feels opportunistic, that's because it probably is.
When Compliance Becomes the Product
The friction isn't abstract. As of March, Twilio's A2P 10DLC policies were still evolving—MMS rate limits now tied to Brand Trust Scores, campaign use cases requiring multi-party approvals, community forums filling up with complaints about rejections and shifting requirements. For teams building agents that need to send appointment reminders or make outbound calls, the gap between "works in dev" and "approved for production traffic" could stretch into weeks. Sometimes longer.
Modi, who previously built AI agent infrastructure at Meta, co-founded AgentPhone with Manav Modi (no relation, apparently), who scaled the Vogue app from roughly 100,000 users to more than a million. Their pitch centers on abstraction: phone number provisioning, they argue, should feel less like navigating telecom regulations and more like infrastructure you barely think about.
"Zero compliance overhead," the website promises. "No carrier bureaucracy."
Whether that holds up under scrutiny—or under the weight of the very telecom regulations they're abstracting away—remains an open question. But for now, developers seem willing to find out.
What You Actually Get
At its core, AgentPhone provides an API for provisioning phone numbers that agents can use to send and receive SMS, make and take voice calls, and support richer messaging protocols down the line. Each agent gets its own number, a model that diverges from traditional Communications Platform as a Service approaches where developers typically manage numbers as shared resources and route messages through application logic.
The platform offers two voice modes. Webhook mode, where developers maintain control over their own LLM and tooling stack. And hosted mode, which includes a built-in LLM for teams that want something running without standing up server infrastructure first.
Voice features include real-time speech-to-text, streaming text-to-speech, barge-in capability, DTMF (touch-tone) support, and backchanneling. Developers can stream live transcripts via Server-Sent Events—useful, one imagines, for monitoring or debugging agent behavior in production, or just watching things go sideways in real time.
All inbound communication—SMS, iMessage, voice transcripts—gets normalized into a unified webhook event model. An inbound message, regardless of channel, triggers an agent.message event. When a call ends, agent.call_ended delivers the full transcript and call analysis. Webhooks are signed with HMAC-SHA256, include replay-window guidance, and follow a six-attempt exponential backoff retry schedule spanning immediate delivery through 12 hours.
For teams that want to let users talk to an agent directly from a browser without provisioning a phone number first, AgentPhone also supports "web calls" via a token-based system and a JavaScript SDK.
It's thorough. Maybe more thorough than the founders initially expected would be necessary.
Developer Ecosystem, Slowly Taking Shape
AgentPhone has SDKs in Python (published to PyPI in early June) and TypeScript/JavaScript. The company also built a Model Context Protocol server, which matters if you're building agents in environments like Cursor or Claude Desktop—making it possible to expose phone capabilities as tools within those workflows.
On GitHub, the company maintains adapters for OpenAI's Agents SDK and CrewAI, alongside sample integrations like "sentry-on-call." The repositories were actively updated through late May and June, suggesting ongoing development beyond the initial launch, though commit frequency alone doesn't tell you much about traction.
Early integrations include Google's Agent Development Kit (confirmed in an April community post), Replit, Sim AI, LangChain, and Alchemy. Sim AI's integration documentation lists 22 operations across numbers, calls, conversations, contacts, and usage—describing AgentPhone as supporting "outbound AI voice calls" with structured transcript data returned.
Whether any of these integrations are seeing serious production use is harder to say.
The Economics of Agent Telephony
AgentPhone's starter tier defaults to up to 10 phone numbers, 5,000 SMS per month, 300 voice minutes per month, a 30-minute maximum call duration, and one concurrent call. The self-serve provisioning cap is 10 numbers; higher limits are available on request. Call recording costs $5 per month as an add-on. An optional denoising feature runs $0.005 per minute. Voice calls are billed per-second with a $0.01 minimum.
Those economics put the product within reach of developers spinning up early-stage agent projects. Pricing details for higher-volume deployments aren't publicly listed, which is typical for infrastructure startups at this stage—you get a sense of the entry point, then everything else happens over email.
A Suddenly Crowded Space
AgentPhone isn't alone. AgentCall, whose terms of service were last updated May 21, also offers a phone number API for AI agents with SMS and voice support. ClawPhone positions itself similarly. Protocol (protocolami.com) takes a different angle, framing itself as an open protocol for "mobile identity for AI agents" with SIM/eSIM provisioning—more identity-layer than immediate developer API, but addressing a related problem.
The broader market for voice agent platforms includes established players like Bland AI and Vapi, though these focus more on call orchestration than assigning individual phone identities to agents. Traditional providers like Twilio, Telnyx, and Vonage remain the underlying infrastructure many teams stitch together—which is precisely the workflow AgentPhone is trying to replace.
The telecom services market was projected to grow from around $2.224 trillion (Grand View Research) to around $2.288 trillion (Fortune Business Insights) in 2026, figures that provide context for the "disrupting a $2 trillion industry" framing that appeared in AgentPhone's Y Combinator launch materials and LinkedIn posts. Whether the company captures a meaningful slice of that depends less on total market size and more on whether developers building autonomous agents find the compliance abstraction compelling enough to adopt a new infrastructure layer.
Big markets are easy to invoke. Capturing them is another matter entirely.
When Agents Need to Pass as Human
Sometime between late May and early June, Manav Modi posted on LinkedIn about an agent successfully ordering pizza using a phone number provisioned through the platform. A small anecdote, maybe. But one that underscores a practical reality: for agents to interact with legacy systems—restaurants, call centers, appointment schedulers—they need phone numbers that look and behave like human phone numbers.
The friction isn't just technical. It's about trust and interoperability with the broader communications ecosystem, systems built over decades with assumptions about who—or what—is on the other end of the line.
AgentPhone's bet is that as more developers build agents capable of making calls, scheduling appointments, or handling customer service, the overhead of managing telecom compliance will become unacceptable. Whether that thesis holds depends on how quickly the agent developer ecosystem grows, and whether abstractions like this one become table stakes or remain niche infrastructure for early adopters.
What Comes Next
For now, the company is a two-person team in San Francisco, part of Y Combinator's Spring 2026 cohort. A LinkedIn post from Zero Index in mid-June mentioned backing the company, though no formal funding round has been disclosed beyond the standard YC deal.
The platform is live. GitHub repositories are active. SDKs are published. Early integrations are in place.
Whether any of this translates into the kind of traction that turns a compliance abstraction into critical infrastructure—that's the part we'll have to wait and see. The problem AgentPhone is solving is real enough. The question is how many developers are far enough along in building AI agents to feel the pain acutely, and how many are willing to bet on a startup to make it go away.
In the meantime, somewhere in San Francisco, an AI agent probably just made a reservation for dinner. And no one had to file paperwork with The Campaign Registry to make it happen.
