# Talk to an app in the browser or on a device

> Talk to an app in the browser or on a device: 9 companies fully do it, 9 with a limit, 29 carry the tag. Every listing in Conversational voice agents with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/voice-agents.web-voice-assistant/
- Markdown: https://www.anchorterminal.com/niches/voice-agents.web-voice-assistant/index.md (~2,050 tokens)
- Slim: https://www.anchorterminal.com/niches/voice-agents.web-voice-assistant/index.min.md (~430 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/voice-agents.web-voice-assistant/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 9 | 9 | 29 | 9 | 4 |

## What counts as done

A short-lived token minted server-side, interruptions honoured, and a client-side tool runs without the model speaking its plan aloud.

Measured, for our own tests: first audio back within 1 second of the user stopping.

How agents ask for it: add a voice assistant to my app; let users talk to the agent from the web page; a companion that reacts to tone

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **AssemblyAI Voice Agent API (AssemblyAI, Inc.)**: Server mints a single-use token with expires_in_seconds and max_session_duration_seconds; the browser connects with ?token=. Client function tools return tool.result, and hold mode keeps the agent from talking during a tool.
- **Deepgram Voice Agent API (Deepgram, Inc.)**: The server mints a short-lived token via /v1/auth/grant with ttl_seconds and the Browser Agent SDK takes a tokenFactory; client-side functions run in the browser and return FunctionCallResponse.
- **ElevenLabs Agents API + MCP (Eleven Labs Inc.)**: GET /v1/convai/conversation/token returns a WebRTC session token and a signed-URL endpoint exists, so the key stays server-side; client tools run in the browser with a configurable response timeout.
- **Inworld Realtime API (Theai, Inc. dba Inworld AI)**: WebRTC with a session token minted by the backend so the Basic key stays server-side, server VAD interruption, and client-run tools with auto continuation.
- **LiveKit Agents (LiveKit Incorporated)**: Frontends join with a short-lived access token minted on the server with the API secret; WebRTC transport, interruptions and client RPC tools are documented.
- **OpenAI Realtime API (OpenAI OpCo, LLC)**: POST /v1/realtime/client_secrets mints an ek_ token lasting 10 seconds to 2 hours (default 10 minutes) for WebRTC in the browser; speed is settable from 0.25 to 1.5.
- **Pipecat (Daily)**: The client SDK calls the developer's /api/start endpoint, which creates a Daily room and returns a room token, so the API key stays on the server; RTVI carries client-side function calls.
- **Retell AI API + MCP (Retell AI Inc.)**: POST /v3/create-web-call returns a call-scoped access_token with expires_at for the browser SDK, so the API key stays on the server.
- **Ultravox Realtime API (Theai, Inc. dba Inworld AI)**: The server creates the call and hands the client only its joinUrl, client tools run in the browser, and the agent reaction setting (listens or speaks-once) stops it narrating around a tool.
- **Vapi API + MCP (Vapi Inc.)**: Web SDK with public keys and short-lived JWTs signed server-side with origin restrictions, plus client-side tools handled in the browser.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Amazon Nova 2 Sonic (Amazon Web Services, Inc.)**: Browser clients reach Sonic through a WebSocket server such as AgentCore with SigV4 auth; there is no short-lived client token in the Sonic docs, so the developer builds the auth layer.
- **Bland AI API + MCP (Intelliga Corporation)**: An embeddable voice and chat widget identified by a public widget_id keeps the key off the page, but there is no token endpoint for a custom client and no client-side tool mechanism.
- **Bolna API + MCP (Voxlabs Private Limited)**: The Web Call SDK takes a backend-minted ephemeral session (about 120 second TTL) so the key stays server-side. Client-side tools are not documented.
- **Gemini Live API (Google LLC)**: Ephemeral tokens from auth_tokens, single use, locked by liveConnectConstraints. Resumption must be granted inside the token constraints or no handle arrives (acc-23), and one tester found replies unnaturally fast (acc-07).
- **GPT-Live (OpenAI OpCo, LLC)**: WebRTC with the server exchanging the browser's SDP offer, so the key stays server-side, but there is no short-lived client token. Client delegation runs tools in the app and can suppress the spoken acknowledgement.
- **Rime Prism (Rime Labs, Inc.)**: No browser tokens or ephemeral keys: the app must stream the microphone through the developer's backend, which holds the key and adds a hop before first audio.
- **Synthflow API + MCP (AgentFlow AI GmbH)**: A backend call returns a per-session media URL and sid for a WebSocket audio stream, so the API key stays server-side, but no client-side tool mechanism is documented and the sid is not described as short-lived.
- **Vaani Developer Platform (Vaani AI Research Private Limited (Bengaluru, India, CIN U62099KA2025PTC201212))**: POST /api/webrtc/session/create returns a session_token (session_ttl 60 to 86,400 s, default 30 minutes) that the client exchanges for WebRTC credentials, so the key stays server-side. No client-side tools documented.
- **Vocily AI (VOCILY AI PRIVATE LIMITED)**: A website widget has voice with a public key that can be rotated, so the API key stays private, but it is a long-lived public key rather than a short-lived token and no client-side tools are documented.
- **Vogent API (Monoid, Inc.)**: GET /dials/{id}/token returns a dial token described as safe to pass to a client, so the key stays server-side, but richer browser flows need advancedBrowserInteraction set with Vogent's help, and no client-side tools are documented.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **Hume EVI (Empathic Voice Interface) (Hume AI, Inc.)**: EVI docs withdrawn (redirect to Audio Decisions), so the access-token flow could not be re-read. acc-27 moved a coding assistant's default voice from Hume EVI 3 to ElevenLabs for naturalness, citing about 300 ms latency.

- not this job: CueLight (sshah03 (individual GitHub account, no name or organisation stated)): No session, token or speech model; a component only. The browser demo runs locally.
- not this job: fone.me (Fone): The website widget lets visitors call the owner's fone.me agent by voice, but it is a keyless embed of that agent, not a session a developer controls with client tools.
- not this job: Fresha AI Concierge (Fresha.com SV Ltd): Phone, SMS and in-app chat for Fresha clients only; no web voice SDK or token.
- not this job: Harmony.ai (monday.com Ltd.): No browser voice SDK or client token in the API; the API covers contacts, workflows, calls and webhooks only.
- not this job: Pipecat PhoneLLM (Daily): No browser session, token or audio; a text model only.
- not this job: PlaceCall (VOYGR Tech, Inc.): Phone calls to businesses only; no browser or app voice session.

## What stops agents finishing

- no browser token, so the key would ship to the page
- ephemeral token runs in restricted mode until resumption is granted server-side
- voice too fast or looping with no pacing control

## Around it

Before: auth.session-token. After: agent.tool-use.

## Where this comes from

Weight 9 of 100 in Conversational voice agents: 4 of 31 accounts (acc-06, acc-07, acc-23, acc-27)

As data: https://www.anchorterminal.com/api/v1/use-cases/voice-agents.web-voice-assistant.json, or the MCP tool get_use_case with id voice-agents.web-voice-assistant. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
