The pitch sounds almost quaint in an era of sprawling chatbots and autonomous code generators: what if your AI agent could just... call someone?
Meet Modi and Manav Modi—brothers, as it happens—think they've spotted a genuine gap in the agentic infrastructure stack. Their company, AgentPhone, emerged from Y Combinator's Spring 2026 cohort with a proposition that borders on the obvious once you hear it. AI agents can scrape websites, write Python scripts, and parse gigabytes of data. But ask one to send a text message or dial a customer, and suddenly you're neck-deep in telephony providers, compliance paperwork, and webhook spaghetti.
AgentPhone offers a way out: a unified API that provisions phone numbers, handles SMS and iMessage, and manages voice calls with real-time transcription. For developers building AI agents that need to operate in the real world—where people still answer phones and read texts—it's infrastructure designed to abstract away the messy parts.
Telephony as a service (with less pain)
The company's core offering is straightforward, perhaps deceptively so. Through a REST API, developers can provision US and Canadian phone numbers and attach them to agents. The numbers themselves come via Twilio, but AgentPhone takes on the compliance burden that often derails engineering teams who just want to ship a feature.
Messaging support covers both SMS and iMessage, including capabilities that feel lifted from consumer apps. Developers can trigger iMessage effects—those "slam" or "gentle" animations that iPhone users either love or tolerate—along with inline replies and multi-image carousels. Inbound messages from both channels funnel into a single webhook, sparing teams from maintaining parallel message handlers.
Voice functionality splits into two modes, each with distinct tradeoffs. Webhook mode streams speech-to-text in real time to a customer's backend for $0.13 per minute, billed per-second. Hosted mode, priced at $0.22 per minute, uses AgentPhone's built-in language model to run autonomous conversations without requiring developers to spin up their own servers. Both deliver full call transcripts when conversations end; developers who need it can stream live transcripts via server-sent events.
It's infrastructure that works, more or less, the way you'd hope infrastructure should work.
The compliance headache nobody wants
Here's where AgentPhone might earn its keep: A2P 10DLC registration. US carriers tightened the screws on commercial SMS earlier this year, requiring brand registration and campaign approval before messages can flow freely. For startups trying to move quickly, it's the sort of bureaucratic friction that can stall a launch for weeks.
AgentPhone handles that registration process. Inbound SMS works immediately upon provisioning a number. Outbound SMS requires the 10DLC setup, though voice calling sidesteps these restrictions entirely. The documentation is refreshingly blunt about what works out of the box and what requires paperwork.
For teams building agents that need to text and call—customer support bots, appointment reminders, verification systems—this matters more than marginal differences in per-minute rates.
Pricing that makes sense (mostly)
Phone numbers cost $3 per month. SMS runs $0.02 per message, whether inbound or outbound. Voice minutes land at $0.13 in webhook mode or $0.22 in hosted mode, both billed per-second rather than rounding up to full minutes. Optional add-ons include call recording ($5 monthly) and noise cancellation ($0.005 per minute).
New accounts get $5 in credit to experiment. Default limits cap self-serve users at 10 numbers, 5,000 SMS per month, 300 voice minutes per month, and a single concurrent call with a 30-minute ceiling. The constraints feel reasonable for a young company managing infrastructure costs while trying to avoid abuse.
Playing well with others
AgentPhone launched with integrations across several agent frameworks—LangChain published an official toolkit available on PyPI, while Sim AI's documentation lists AgentPhone as a supported tool for dynamic number provisioning and two-way messaging. Google's Agent Development Kit has a dedicated integration page emphasizing "no Twilio account, no ngrok, no server needed," which is the kind of promise that resonates with developers who've burned hours on webhook debugging.
The company also ships a Model Context Protocol server compatible with Cursor, Claude Desktop, and Windsurf, distributed as an npm package. MCP tools include operations like buy_number, make_call, and make_conversation_call—the basics, done cleanly.
According to the company's launch announcement, teams at Google ADK, Replit, Y Combinator, Sim AI, LangChain, and Alchemy were early adopters. Documentation confirms the Sim AI and LangChain integrations; the Google ADK integration has a published page; other partnerships remain unconfirmed publicly, which is either cautious partnership messaging or aspirational claims getting ahead of reality.
What comes next
The initial release covers SMS, iMessage, and voice calling over traditional phone networks. WhatsApp and RCS are planned additions, though neither has surfaced in the public API documentation yet. Perhaps the team is navigating Meta's developer policies, or maybe they're simply focused on nailing the core product first.
The founders
Meet Modi previously worked at Meta, where he built agent infrastructure on WhatsApp—a platform that, according to the company, serves over 280 million businesses. Manav Modi comes from consumer growth roles, including work on Vogue's app. The contrast is useful: one brother understands enterprise infrastructure at scale, the other knows how to make consumer products people actually use.
The team is lean—just the two of them, according to their Y Combinator profile. They hosted a "Call My Agent" hackathon at Y Combinator's headquarters in May, the kind of developer outreach that signals a team interested in community feedback rather than just pushing API docs into the void.
Why this still matters
Phone numbers occupy a strange position in modern software infrastructure. They're old technology—unreliable, expensive, tangled in carrier politics—yet they persist as trusted identifiers. Industry analysts have noted that only a fraction of agent interactions will be fully automated in the near term, leaving plenty of room for hybrid systems where AI handles routing but humans still need to get involved.
A recent analysis from DataDome flagged an "AI Agent Identity Crisis," noting that most autonomous agents lack robust authentication signals. Phone numbers, for all their flaws, carry weight as trusted identities in systems built for human interaction. AgentPhone's bet is that agents will need to operate in spaces where business still happens the old-fashioned way: through calls and text messages.
The infrastructure is live. Pricing is transparent, documentation is public, and the integrations suggest real adoption beyond launch-day hype. For developers building agents that need to interact over phone and text, the stack is, as promised, one API call away. Whether that's enough to build a durable business remains to be seen—but at least the phones are ringing.
