AgentPhone

AgentPhone provisions US/Canada phone numbers and routes both messaging (SMS, iMessage, WhatsApp, MMS) and voice calls through a single webhook format, so an agent's backend doesn't need separate code paths for texting and calling. Voice can run in 'webhook' mode (your server generates every response) or 'hosted' mode (AgentPhone's own LLM handles the call end-to-end from a system prompt, no server required). Both modes share the same engine: streaming speech-to-text, streaming text-to-speech, caller barge-in, DTMF keypad input for IVR navigation, call transfer, and voicemail. Outbound calls can also be placed with a one-shot 'make_conversation_call' that needs no webhook at all — you just pass a system prompt and AgentPhone runs the conversation. Source: https://docs.agentphone.ai/documentation/guides/calls and https://docs.agentphone.ai/welcome

The FAQ is explicit about the comparison: Twilio gives low-level telecom primitives, AgentPhone is an opinionated layer built specifically for agents on top of that. Concretely: one webhook shape covers both calls and texts instead of separate pipelines; conversation threading and recent-message context are handled for you; real-time NDJSON streaming lets TTS start speaking on the first token instead of waiting for a full LLM response, which matters a lot for perceived voice latency; and a live SSE transcript stream lets you monitor an in-progress call without polling. It also natively supports iMessage and WhatsApp alongside SMS, which most 'agent phone number' competitors don't bother with. Source: https://docs.agentphone.ai/documentation/guides/calls, https://docs.agentphone.ai/welcome

Pay-as-you-go, no subscription tiers. Every account starts with $5.00 free credit and no credit card required. Published rates as of 2026-08-28: phone numbers $3.00/month each, SMS $0.02/message (in or out), voice in webhook mode $0.13/min, voice in hosted (built-in AI) mode $0.22/min, both billed per second with a 1¢ minimum, and an optional call-recording add-on at a flat $5.00/month. Phone number lookup (line-type / RCS-capability check) is $0.009 per number. Self-serve accounts can hold up to 10 numbers before needing to contact the vendor for a higher limit; provisioning a number requires at least $3.00 balance. Source: https://docs.agentphone.ai/usage

Inbound SMS and all voice calling (inbound and outbound) work immediately after provisioning a number. Outbound SMS is different: US carriers require 10DLC (10-Digit Long Code) registration before a number can send texts, which AgentPhone handles by having you fill out a form (vendor says ~5 minutes) and submitting it to the carriers on your behalf, or booking a call for hands-on help. This is a US-carrier requirement that would apply to any legitimate SMS platform, not something AgentPhone imposes — but it means 'send an SMS' isn't instantly available the way voice is, and a buyer building a texting-heavy workflow should budget for that registration lag before launch. Source: https://docs.agentphone.ai/documentation/guides/phone-numbers

AgentPhone is Y Combinator Spring 2026 (batch P26), founded by brothers Manav Modi (CEO) and Meet Modi (CTO); the YC company page lists a team of 2, though they have an open Founding Engineer req ($140K-$250K + 0.25-2% equity). Meet's bio states he 'previously built AI agent infrastructure on WhatsApp for 280+ million businesses at Meta' and Manav's bio states he 'grew usage from ~100k to over 1M users' on the Vogue app — both are founder-supplied bio claims we could not independently verify. The company's YC launch post (May 2026) names Google's Agent Development Kit, Replit, Sim AI, LangChain, Vercel, and Alchemy as teams 'building with AgentPhone'; this is corroborated in the sense that each of those platforms hosts its own real integration/adapter docs page for AgentPhone (e.g. docs.langchain.com, adk.dev, chat-sdk.dev), but that confirms integration exists, not paying usage at any particular scale. We found no independent press, reviews, or case studies beyond YC's own launch coverage and syndicated 'YC batch' roundup articles. Source: https://www.ycombinator.com/companies/agentphone, https://www.ycombinator.com/launches/QNE-agentphone-phone-numbers-for-ai-agents

AgentPhone's product itself is closed-source (hosted API); what's open-source are client integrations. The MCP server (agentphone-mcp, TypeScript, MIT license) has 118 GitHub stars, 17 forks, and 3 open issues, and was last updated August 27, 2026 — actively maintained. The official Python SDK is on PyPI as 'agentphone', currently at version 0.18.0 (MIT license), maintained by the same founder account. There's no GitHub Releases history on the MCP repo (empty releases list), so there's no formal changelog beyond commit activity. Separate client-library repos exist for CrewAI and the OpenAI Agents SDK, both thin integration wrappers rather than the product itself. Source: https://github.com/AgentPhone-AI/agentphone-mcp, https://pypi.org/pypi/agentphone/json

No security/compliance/trust page: the site links to /trust from its footer route list but that URL currently renders the same generic homepage shell with no distinct compliance content, and no SOC 2, DPA, or sub-processor list is published anywhere we could find. There's no public status page. Security practices that do exist are documented at the developer level only — HMAC webhook signature verification with timestamp checks against replay attacks — which is a reasonable baseline but not a substitute for a compliance page for buyers who need one. None of this is unusual for a two-founder, months-old startup, but it means due diligence on data handling currently requires asking the vendor directly rather than reading a public page.