Retell AI

An Apps framework holds credentials for each external system, with documented agent functions across CRM (Salesforce, HubSpot, Dynamics 365, Zoho, GoHighLevel), helpdesk (Zendesk), scheduling (Calendly, Cal.com) and file storage (Google Drive, OneDrive), plus custom functions calling any HTTP API mid-call, a code tool running JavaScript, and MCP tools and nodes reaching remote MCP servers during a live conversation. Contact data also syncs both ways with the CRM.

Sourcedocs.retellai.com/integrations/overviewread 2026-09-05

Conversation flow agents are node-and-edge graphs with conversation, subagent, function, code, logic split, extract-variable, SMS, transfer and global nodes, transitions gated by either LLM-evaluated prompt conditions or deterministic equations, reusable subflows shared across flows, and Flex Mode compiling the graph into a runtime prompt for varied caller behavior. Multi-prompt state agents and agent-to-agent transfer cover the same ground for prompt-built agents, with versions and environment tags controlling what is live.

Sourcedocs.retellai.com/build/conversation-flow/overviewread 2026-09-05

Knowledge bases are built and maintained by the platform: sources attached as URLs, documents across roughly twenty file formats, or custom text, then chunked, embedded and stored in a vector database at creation time, with retrieval tuned by query-rewrite instructions that condense recent dialogue into a standalone search query. Google Drive and OneDrive sources re-sync on a 24-hour refresh cycle, and knowledge bases are managed independently of any one agent through their own API.

Sourcedocs.retellai.com/build/knowledge-baseread 2026-09-05

Guardrails detect prohibited topics in both agent responses and user input and replace flagged content with a safe placeholder before it is spoken, and live monitoring lets a person follow the transcript in real time, listen in silently, take over from the agent mid-call or end the call, with an Update Live Call endpoint to inject context or trigger a response on a running session.

Sourcedocs.retellai.com/features/live-monitoringread 2026-09-05

SOC 2 Type 1 and Type 2 certified, HIPAA and GDPR compliant with a BAA, DPA and a documented GDPR erasure process, alongside customer-facing controls: role-based access control with Admin, Developer, Analyst and Viewer roles, API keys scoped read or edit per area, a webhook signing key with signature verification and IP allowlisting, expiring signed URLs for recordings and logs, per-agent data storage settings and per-agent retention policy.

Sourcedocs.retellai.com/general/complianceread 2026-09-05

A monitor-call WebSocket streams transcript, tool call and node events from a live session, so the path the agent took through the flow is reconstructable turn by turn, backed by session history with transcripts and CSV export, per-call P50, P90 and P99 latency, disconnection reason codes, call_started, call_ended and call_analyzed webhooks, and AI QA scoring calls on hallucination rate, knowledge base accuracy, tool usage, latency and sentiment against customer-defined cohorts.

Sourcedocs.retellai.com/api-references/monitor-call-websocketread 2026-09-05

Contact records keyed by phone number carry custom fields populated by post-call extraction mappings, so the agent recalls preferences, past issues and commitments across every later conversation, with a merged timeline of that contact's calls and chats and a backfill job that re-applies mappings to historical calls. Scope is per contact and the record is separately deletable. It is implemented as contact field mappings rather than a dedicated memory store with its own lifetime, so retention follows the contact record.

Sourcedocs.retellai.com/integrations/build-contact-memoryread 2026-09-05

The documentation covers what is stored and for how long through per-agent data storage settings, a per-agent retention policy and expiring signed URLs, but those are privacy controls rather than control over where data lives. No region selection, VPC, single tenant, self hosted or on premises option is documented, and the reliability page describes shared enterprise infrastructure without naming a region.

Sourcedocs.retellai.com/accounts/data-retentionread 2026-09-05

The quick start has the customer pick a template as the first build step, Agent Handbook ships toggleable preset prompt packs for personality, accuracy and safety, and reusable subflows package parts of a conversation flow as shared components adopted across agents. The template catalog itself sits inside the dashboard rather than on a public page, so its breadth is not published.

Sourcedocs.retellai.com/get-started/quick-startread 2026-09-05

Inbound and outbound calling on Retell-managed numbers or imported ones over elastic SIP trunking, with documented setup for Twilio, Telnyx, Vonage, Avaya Aura, Genesys Cloud, Five9 and Amazon Connect, alongside browser web calls, a website widget in chat, voice and callback modes, two-way SMS and MMS, scheduled batch call campaigns, and an inbound call and SMS webhook that picks the agent and passes per-call context at ring time.

Sourcedocs.retellai.com/deploy/custom-telephonyread 2026-09-05

The customer picks the language model in agent settings and tunes temperature, fast tier, structured output, retries and timeouts, choosing across providers including OpenAI, Anthropic and Google as named in the model deprecation notices, or runs their own model entirely through the custom LLM WebSocket protocol while Retell handles telephony, transcription and turn-taking. Speech recognition and text-to-speech providers are separately selectable.

Sourcedocs.retellai.com/build/llm-optionsread 2026-09-05

A REST API at api.retellai.com published with an OpenAPI specification and a Postman collection, covering calls, chats, agents, response engines, knowledge bases, contacts, apps, tests and phone numbers; official TypeScript and Python SDKs plus a browser Web SDK; a command line interface; a hosted MCP server at mcp.retellai.com for Cursor, Claude Desktop and Claude Code; signed webhooks; and a documented LLM WebSocket protocol for bringing your own model. A published deprecation feed with dated breaking-change notices sits alongside it.

Sourcedocs.retellai.com/api-references/overviewread 2026-09-05

Five shipped testing surfaces with readable results: an LLM Playground that inspects tool calls and transitions and replays any turn, LLM simulation testing where an AI-simulated user runs a scenario and success criteria grade each run, batch tests over many cases reporting pass rate and per-case results in a testing history, web and phone call testing, and production A/B testing that splits live traffic between agent versions. Test case definitions, batch tests and test runs each have their own API.

Sourcedocs.retellai.com/test/test-overviewread 2026-09-05

Browser, desktop session and remote computer control are all absent from the documentation. The Press Digit tool and press digit node do let the agent infer and press keypad digits to walk a third party's DTMF input IVR menu on an outbound call, and voicemail handling detects and responds to another system's automated flow; that is real and documented, but it works a telephone keypad rather than a browser, desktop or remote computer.

Sourcedocs.retellai.com/build/single-multi-prompt/press-digitread 2026-09-05