Telnyx

Assistants call outside systems with webhook tools, which can run asynchronously with filler messages while the request completes, attach MCP servers for whole tool catalogs, and connect enterprise integrations for Salesforce, Zendesk and Jira. Call control tools transfer the caller, hand off with SIP Refer, send DTMF tones to another phone system and hang up, client side tools run actions in the web widget, and keys for outside providers are kept as write only integration secrets. Tools live in a shared library that several assistants and workflow nodes can reuse.

Sourcedevelopers.telnyx.com/docs/inference/ai-assistants/no-code-voice-assistant.mdread 2026-10-11

Conversation workflows store a directed graph on the assistant. Prompt nodes run the model with their own instructions and optional model and voice overrides, speak nodes play a fixed message, and tool nodes run one shared tool and advance.

Edges move between nodes on a natural language condition the model judges, a deterministic comparison against variables, or a default, and a tool node's HTTP status code can steer the next edge.

An edge can also hand the conversation to another assistant, so a triage assistant can route billing, technical and sales callers to specialists that share the conversation's context.

Sourcedevelopers.telnyx.com/docs/inference/ai-assistants/workflows.mdread 2026-10-11

Each assistant can use named knowledge bases built from uploaded files or a URL, kept with the account and attached in the assistant's configuration, so one knowledge base can serve several assistants. Retrieval during a call is part of the voice engine's orchestration and included in the per minute price, while the price estimator carries a separate agent knowledge base add on. Telnyx does not say how the content is chunked or indexed.

Sourcetelnyx.com/pricing/conversational-airead 2026-10-11

Transfer and SIP Refer tools pass a caller to a person or another phone system, with voicemail detection on transfer, and handoff tools move a conversation between assistants. Releases stay under the team's control. A new version is saved with notes, traffic rules send selected callers or a percentage to it, and rollback returns all traffic to the main version at once. No step lets a person approve a tool call or a reply before the assistant acts.

Sourcedevelopers.telnyx.com/docs/inference/ai-assistants/version-testing-traffic-distribution.mdread 2026-10-11

Business accounts sign in to the Mission Control Portal through SAML single sign on with Okta, Azure Active Directory, Google, OneLogin, Auth0 or LastPass. Telnyx holds ISO 27001:2022 and ISO 27701, a SOC 2 Type II report under NDA covering voice, messaging, video and wireless, and HIPAA coverage for voice, messaging, fax and AI inference with a BAA available, plus PCI DSS for its own billing. The SOC 2 scope does not name AI assistants, and Telnyx also claims a SOC 2 Type III, a report type the AICPA does not issue.

Sourcetelnyx.com/ai/compliance.mdread 2026-10-11

Every conversation is stored and listable through the conversations API, transcripts show which workflow node produced each assistant message, and the last tool call's HTTP status is kept as a variable. Insight groups run structured and unstructured analysis on each conversation and can post results to a webhook, with Telnyx managed insights alongside custom ones. Version notes record what changed in each release, and Telnyx does not mention an audit trail of who changed an assistant's configuration.

Sourcedevelopers.telnyx.com/docs/inference/ai-assistants/workflows.mdread 2026-10-11

Memory brings earlier conversations into a new one. The dynamic variables webhook returns a query in the same form as the List Conversations endpoint, typically the caller's phone number, a time window or custom metadata, plus the IDs of insights to include, and the assistant starts with those conversations and insight results in context, across voice and messaging.

The webhook has one second to answer before the call carries on without memory. What comes back is stored conversation history picked by the customer's query. There is no separate memory store, and Telnyx sets out no retention period or way to delete one person's history.

Sourcedevelopers.telnyx.com/docs/inference/ai-assistants/memoryread 2026-10-11

Voice AI agents run in the United States, the primary data center, or in the European Union with EU data residency, where data does not leave the EU. AI inference is also offered in Asia Pacific with regional processing, though voice agents are not offered there, and Canada has neither. The assistant's language model can also point at the customer's own OpenAI compatible endpoint, such as one on AWS Bedrock, Azure OpenAI or a self run vLLM server, which keeps inference inside an environment the customer controls.

Sourcetelnyx.com/ai/compliance.mdread 2026-10-11

New assistants start from a template in the portal, with a blank template as the usual starting point, and Telnyx managed insights come ready to attach to any assistant. A shared tools library lets one set of tools serve many assistants. Telnyx offers no catalog of finished agents for support, sales or scheduling.

Sourcedevelopers.telnyx.com/docs/inference/ai-assistants/no-code-voice-assistant.mdread 2026-10-11

Calls reach an assistant on an assigned Telnyx number or over SIP, outbound calls start from the portal, the API or the TeXML ai_calls endpoint, and an embeddable widget carries voice and chat on a web page. SMS conversations run through the assistant SMS chat API, with MMS during a call when messaging is on. At the start of each conversation a dynamic variables webhook can load caller details and memory settings, and custom SIP headers or the outbound request can pass variables in.

Sourcedevelopers.telnyx.com/docs/inference/ai-assistants/no-code-voice-assistant.mdread 2026-10-11

The customer picks the model for each assistant on the Agent tab. Options include models hosted on Telnyx GPUs, with Kimi as the default, managed third party models from providers such as OpenAI, Gemini and Groq, with OpenAI models on the customer's own key, and any public OpenAI compatible endpoint such as AWS Bedrock, Azure OpenAI, Baseten, vLLM or SGLang. Prompt nodes in a workflow can override the model step by step, and speech to text and text to speech providers are chosen separately.

Sourcedevelopers.telnyx.com/docs/inference/ai-assistants/custom-llm.mdread 2026-10-11

The API covers assistants end to end, including creating and updating assistants, assistant tests, canary deploys, MCP server records, SMS chat and a chat endpoint in beta, and the TeXML ai_calls endpoint starts outbound calls. Coding assistants can read the whole reference through llms.txt indexes, and the conversations API hands stored conversations to other systems.

Sourcedevelopers.telnyx.com/development/llms/ai-assistants-llms-txtread 2026-10-11

AI tests give each check a name, a target assistant and success criteria, such as what the greeting must say, and a test run shows progress live and reports whether the assistant met every criterion, with the conversation to review. A new version can be saved with notes and sent a share of traffic by ordered rules on the caller's number, all to one version or split by percentage, with the remainder staying on the main version, and rollback clears every rule at once. Coval can generate test conversations from seed cases.

Sourcedevelopers.telnyx.com/docs/inference/ai-assistants/version-testing-traffic-distribution.mdread 2026-10-11

Assistants act through webhooks, MCP servers, integrations and call control tools, and send keypad tones to other phone systems. No hosted browser, remote desktop or screen control is offered.

Sourcedevelopers.telnyx.com/docs/inference/ai-assistants/tools-library.mdread 2026-10-11