The pitch is almost comically straightforward. You're building a conversational AI agent, you need it to make and receive calls, and you'd rather not spend three days wrestling with Twilio's telecom primitives just to route a simple conversation.
That's the bet AgentPhone is making. The San Francisco startup—two brothers, fresh out of Y Combinator, with no publicly announced funding beyond the accelerator's standard check—launched an API this month designed to collapse what used to take developers dozens of integrations into a single webhook. Voice, SMS, even iMessage. One endpoint handles it all, they say, with real-time transcription and conversation threading baked in from the start.
It's infrastructure purpose-built for the age of AI agents, not retrofitted from the interactive voice response systems that dominated call centers a decade ago. Whether that distinction matters enough to carve out space in a $2 trillion global telecom industry is the question Manav and Meet Modi are now trying to answer in real time.
Simplicity as Strategy
The Modi brothers call their approach "agent-native," which in practice means aggressive abstraction. Where Twilio hands developers low-level building blocks—SIP trunks, media streams, codec negotiations—AgentPhone offers something closer to what an AI engineer actually wants: a phone number that just works.
Provision a US or Canadian number with a single POST request, optionally specifying an area code. Inbound calls, outbound calls, SMS messages—everything funnels into one webhook. The API spits back real-time transcripts, threaded conversations, post-call summaries. No stitching together services. No decoder ring required.
The platform supports two voice modes, and the split reveals the target market. In "webhook" mode, developers receive transcribed conversation turns as agent.message events and stream responses back as NDJSON—full control, maximum flexibility. In "hosted" mode, AgentPhone's own large language model runs the entire call using just a system prompt, handling speech-to-text, text-to-speech, barge-in interruptions, and touchtone detection server-side. That second option? It's for teams that want agents on the phone by Friday, expertise be damned.
Pricing runs $3 monthly per number, 13 cents per minute for webhook voice, 22 cents for hosted voice, 2 cents per SMS. The starter tier caps users at 10 numbers, 5,000 messages, and 300 voice minutes a month.
Betting on the Model Context Protocol
One early bet stands out. AgentPhone shipped with support for the Model Context Protocol—MCP—right out of the gate. It's a technical detail that matters more than it might seem.
Developers using Claude Code, Cursor, Windsurf, or other MCP-compatible coding environments can now integrate AgentPhone directly into their workflow through the agentphone-mcp server. Python and Node SDKs went live in mid-May, with the Python package already cycling through updates on PyPI. GitHub shows integration examples for CrewAI and OpenAI's Agents SDK. There's even a published Claude Code skill for quick setup.
Third-party adoption is starting to surface. Sim.ai, an agent platform, documented over 22 AgentPhone operations for managing numbers, calls, conversations, and contacts. The Agent Development Kit marketplace now lists AgentPhone as a telephony tool. It's early, perhaps too early to call it momentum, but the connective tissue is forming.
Who They Are
The brothers splitting this workload come with relevant but not identical backgrounds. Manav previously worked on Vogue's app revamp, reportedly growing usage from around 100,000 users to over a million—consumer mobile, brand-driven growth. Meet built AI agent infrastructure on WhatsApp at Meta. Telecom-adjacent experience, but not pure telecom.
They came through Y Combinator's recent batch with David Lieb as their primary partner. No seed round announced. No disclosed venture backing beyond YC's standard investment. They're hosting a "Call My Agent" hackathon on May 17 at YC's San Francisco office, positioning it as a sandbox for developers to experiment with agent telephony.
The setup is lean by design or necessity, hard to say which.
The Infrastructure Underneath
Peek under the hood and the startup's strategy becomes clearer. Documentation indicates AgentPhone provisions phone numbers through Twilio while routing SMS delivery through a downstream carrier called Sinch. In other words, they're layering abstraction over existing telecom infrastructure, not building carrier relationships from scratch—at least not yet.
That hybrid approach makes sense for a two-person team trying to move quickly. It also raises questions about margins, control, and what happens when Twilio or Sinch changes pricing or terms. For now, the platform supports A2P 10DLC messaging with detailed rate limits and queueing for over-limit sends. iMessage support includes daily caps for different message types, plus reactions, though the technical implementation remains undisclosed.
Coverage is limited to the US and Canada, with the website offering the standard "more countries coming soon" disclaimer.
A Suddenly Crowded Space
The timing is both opportune and perilous. Incumbents are scrambling to retrofit themselves for what everyone's now calling the "agentic era."
Twilio announced a "next-generation platform" in early May, explicitly positioning it as infrastructure for AI agents and releasing real-time Conversation Intelligence around the same time. Telnyx launched something called ClawdTalk in February for agent frameworks, then followed with LiveKit on Telnyx in April for deploying voice AI agents. Dialpad shipped an Agentic AI Platform last fall.
Newer players are proliferating faster than anyone can track. Retell AI lists voice agent pricing between 7 and 31 cents per minute. Vapi runs a widely-used voice agent platform on usage-based minutes. AgentCall charges a $19.99 monthly fee with AI voice minutes at 40 cents. Ring-a-Ding launched an OpenClaw skill for outbound agent calls in April.
The naming conventions alone reveal the moment we're in: multiple services now use variations of "agent" plus "phone" or "number" in their branding, some entirely unrelated to the YC company. It's the sort of convergent evolution that happens when an entire industry spots the same opening at once.
What Developers Are Actually Saying
Scroll through developer forums from recent months and a clear frustration emerges. Codec renegotiation headaches. Barge-in implementation challenges. Silent disconnects that require telecom expertise to debug. The recurring theme: building conversational AI shouldn't demand mastery of telephony's arcane internals.
AgentPhone's bet—and it is a bet—is that the market wants simplicity over flexibility. One unified webhook instead of separate integrations for voice and messaging. Automatic transcription instead of manual speech-to-text wiring. MCP support instead of custom tool implementations.
Whether that abstraction layer holds up under production scale, edge cases, and the regulatory complexity that defines global telecom will determine if two people can carve out durable territory in an industry measured in trillions. Abstraction is elegant until it breaks. Then it's just another layer to debug.
The API is live, the SDKs shipped mid-May, and developers can provision numbers starting today. What happens next—whether AgentPhone becomes infrastructure or a footnote—depends on how many teams decide telecom is someone else's problem worth paying to solve.
