AgentPhone

Hosted agents act only on the phone line itself: they can cold transfer a caller to a set number, press keypad digits to get through a phone menu, and read or send SMS and iMessage during a call. No function calling, custom tools or connector catalog exists for hosted agents, so a lookup in a CRM or calendar happens only when the customer's own webhook server runs it and streams the answer back. The channels themselves are a fixed set: voice, SMS, MMS, iMessage, WhatsApp and RCS.

Sourcedocs.agentphone.ai/documentation/guides/callsread 2026-11-09

A call follows one fixed path: streaming speech recognition, a reply from the hosted model or the customer's webhook, then streaming speech synthesis, with barge in, backchannel cues and turn detection handled by the platform. The only handoffs are a cold transfer to a phone number and a text sent during the call. There is no flow builder, branching logic or way to pass a caller between several agents, and each phone number belongs to exactly one agent, so multistep routing lives behind the webhook in the customer's own code.

Sourcedocs.agentphone.ai/documentation/guides/callsread 2026-11-09

Hosted agents work from a system prompt, a greeting and template variables passed in when a web call starts, and nothing else. The platform has no document store, knowledge base, file upload or retrieval index, so an agent that must answer from a company's policies, product catalog or customer records needs that lookup to happen in the customer's webhook backend, which receives each transcribed turn and can query whatever source the team uses before replying in text.

Sourcedocs.agentphone.ai/documentation/guides/agentsread 2026-11-09

People stay in the loop through limits rather than approvals. An agent can hand a caller to a person by cold transfer to a configured number, and a hosted agent makes that call itself when someone asks for a human. The MCP purchase tool refuses to buy a number until it is called again with an explicit confirmation, texting arbitrary numbers stays off unless the account is cleared for it, and iMessage sending follows daily quotas and reengagement limits. No queue holds a call or message for review before it goes out.

Sourcedocs.agentphone.ai/documentation/reference/faqread 2026-11-09

Access runs on API keys created in project settings and sent as bearer tokens, and sub accounts isolate the numbers, agents and conversations of separate brands or clients under one parent key. Customer staff get no named team roles, permission model or single sign on. AgentPhone completed a SOC 2 Type 1 examination by an independent CPA firm covering security, availability and confidentiality, with Type 2 underway and the report shared under NDA, encrypts data with AES 256 at rest and in transit, and requires multifactor authentication for its own staff.

Sourceagentphone.ai/trustread 2026-11-09

Each call keeps a record with direction, status and duration, a turn by turn transcript pairing what the caller said with the agent's reply and a confidence score, and a recording when the $5 add on is active. Transcripts stream live over server sent events or replay after the call, and the call ended event carries a summary. Webhook delivery logs and statistics, plus usage by number, agent, day or month, round out the view. Transfers, keypad presses and texts sent mid call do not show up as separate actions in the record.

Sourcedocs.agentphone.ai/documentation/guides/callsread 2026-11-09

Every exchange with a contact lands in a threaded conversation that persists across calls and messages, and each webhook carries recent history for context. A conversationState field lets the customer's backend store its own metadata on a thread, which then rides along in every later payload so the agent picks up where it left off without a separate database. Conversation data stays while the account is active and is deleted on request. The agent writes nothing to memory on its own, and items carry no retention setting.

Sourcedocs.agentphone.ai/documentation/reference/faqread 2026-11-09

AgentPhone runs as a hosted service from the United States, and data from the EEA, the UK and Switzerland moves there under standard contractual clauses rather than staying in a region the customer picks. No region selector, private cloud or self hosted option exists. Numbers come from the US and Canada, with UK SMS only numbers on enabled accounts, and SIP trunking can route calls into the customer's own voice platform, which changes where calls are answered but not where AgentPhone keeps transcripts and messages.

Sourceagentphone.ai/privacyread 2026-11-09

AgentPhone ships no prebuilt agents, call scripts or templates for receptionists, sales follow up or reminders. A team writes each hosted agent's system prompt and greeting itself or builds the logic behind the webhook. What comes ready made is integration code: a Claude Code skill, a skills file that OpenClaw and Hermes agents follow to get a number, adapters for Vercel's Chat SDK, LangChain and Google ADK, and code examples for webhook handlers, voice tool calls and browser calls.

Sourcedocs.agentphone.ai/llms.txtread 2026-11-09

Inbound calls and texts start work on their own. A call to an attached number is answered by the hosted agent or posted to the customer's webhook turn by turn, and every inbound SMS, MMS, iMessage, WhatsApp or RCS message fires an agent.message event carrying recent history. Call endings fire agent.call_ended with the full transcript and a summary, iMessage reactions fire their own event, and webhooks can be set per agent or for the whole project. Outbound calls start from an API request, and browser web calls connect without any phone number.

Sourcedocs.agentphone.ai/documentation/reference/faqread 2026-11-09

In webhook mode, the default, the customer's own server writes every reply, so a team runs any model, provider, retrieval setup or custom logic it chooses and can switch without touching AgentPhone. Hosted mode instead runs a built in model picked by tier: turbo for the lowest latency, balanced as the default, or max for complex reasoning, and AgentPhone does not name the provider behind those tiers. Speech recognition can be set to fast or accurate, and voices come from a catalog listed through the API, with a per call voice override.

Sourcedocs.agentphone.ai/documentation/guides/agentsread 2026-11-09

A REST API with an OpenAPI 3.1 file covers agents, numbers, messages, conversations, contacts, calls, webhooks, usage, verification, WhatsApp templates and sub accounts, alongside official Python and TypeScript SDKs and a browser SDK for web calls. The MCP server runs hosted over streamable HTTP with OAuth sign in or a bearer API key, or locally through npm, and exposes 28 tools that buy numbers, create, update and delete agents, attach numbers, place calls, send messages and manage webhooks and contacts.

Sourcedocs.agentphone.ai/mcpread 2026-11-09

Testing covers the plumbing rather than the agent's answers. A test endpoint sends signed sample events to a webhook so a team can check delivery and signature handling, browser web calls let a developer talk to an agent without a phone line, and recordings and transcripts let a team listen back after changing a voice setting or prompt. There are no scored test cases, simulated callers, evaluation sets or comparisons between agent versions, so judging whether a new prompt answers better is left to the team.

Sourcedocs.agentphone.ai/documentation/reference/testingread 2026-11-09

Agents on AgentPhone never drive a browser, desktop or remote computer. The closest behavior is phone menu navigation, where an agent presses keypad digits to work through an automated phone system, and hosted agents do this unprompted when a menu asks for input. Everything else runs through programmatic interfaces: the REST API, webhooks, the MCP server and carrier networks for calls, SMS, iMessage, WhatsApp and RCS, with browser web calls used only for a person to talk to the agent.

Sourcedocs.agentphone.ai/documentation/reference/faqread 2026-11-09