Niches · Conversational voice agents
Run a tool during a live call and tell the caller the true outcome
Two or three companies' docs show a full fit.
What counts as done
Precisely one execution per intended action, a spoken confirmation that matches the tool's real result, and the conversation resumes without a client nudge in one response cycle.
Measured, for our own tests: no more than 4 seconds of silence before a filler or result. No docs we read state these figures, so they don't count against a listing here.
How agents ask for it: book the slot while the caller is on the line
, look up the order and read it back
, issue the refund and confirm it
Does the job 2
The docs show it meets the test above, and nothing we read says otherwise.
| Listing | Why |
|---|---|
| GPT-Live BB OpenAI OpCo, LLC | Backend work keeps running while GPT-Live talks; an automatic acknowledgement covers the wait (acknowledgement_delay_ms), and session.commentary.append speaks the verified result without a new response request. Appended instructions do not cancel backend work. source |
| Ultravox Realtime API C Theai, Inc. dba Inworld AI | Non-precomputable tools are invoked only after the agent's utterance ends and the conversation is frozen while they run, so barge-in does not cancel a write; the agent speaks after the result by default (agent reaction), timeout 2.5 s default, 40 s max. source |
Does the job, with a limit 19
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.
| Listing | Why |
|---|---|
| Amazon Nova 2 Sonic B Amazon Web Services, Inc. | Tools run asynchronously and are not cancelled by new speech; toolChoice can force a tool. No cancellation event, results are always delivered so the model may issue a second call, and a LiveKit account saw results held until the user spoke (fixed). source |
| AssemblyAI Voice Agent API C AssemblyAI, Inc. | Per-tool execution_mode: interactive speaks a filler while the tool runs, hold mutes replies; tool.result auto-fires the next reply. What happens to an in-flight tool on barge-in, any cancellation event, and tool forcing are not documented. source |
| Bland AI API + MCP B Intelliga Corporation | Agent speech fills the wait and a 1 to 30 second cooldown limits repeat calls, but nothing documents whether caller speech cancels a running tool or whether a cancelled tool is re-issued. source |
| Bolna API + MCP D Voxlabs Private Limited | pre_call_message fills the wait while an API tool runs. No tool timeout range, cancellation-on-interruption behaviour or dedupe of re-issued calls is documented. source |
| Deepgram Voice Agent API BB Deepgram, Inc. | FunctionCallCancelled tells the client which calls were abandoned and late responses are dropped; defer_until_eot stops speculative dispatch. But barge-in still cancels a running call with no per-tool opt-out, and tool timeouts are undocumented. source |
| ElevenLabs Agents API + MCP BB Eleven Labs Inc. | Per tool: pre_tool_speech auto, force or off; execution_mode immediate, post_tool_speech or async; interruption_mode can block barge-in during the tool; timeouts 5 to 300 s (webhook) or 1 to 120 s (client). Errors are hidden by default and no cancellation event is documented. source |
| Fresha AI Concierge F Fresha.com SV Ltd | Books against the Fresha calendar during the call per the vendor, but no developer tools can be registered and nothing documents fillers, timeouts or what the caller hears if a booking fails. source |
| Gemini Live API C Google LLC | Per-function BLOCKING or NON_BLOCKING behaviour, INTERRUPT, WHEN_IDLE or SILENT scheduling and a toolCallCancellation message are documented, but a user test found no cancellation and the same booking re-issued under a new id in 7 of 12 runs. source |
| Inworld Realtime API B Theai, Inc. dba Inworld AI | auto_tool_response (default true) continues after a function_call_output, responsiveness fillers cover slow tools and tool_choice can force a named function. The docs don't say what happens to a pending call when interrupt_response cancels the response. source |
| LiveKit Agents BB LiveKit Incorporated | Per-tool disallow_interruptions, hold messages, tool_choice per reply and automatic continuation are documented, but on_duplicate defaults to allow and matches duplicates by tool name only, so precisely-once needs explicit settings. source |
| OpenAI Realtime API BB OpenAI OpCo, LLC | Function tools run in the client and tool_choice works at session and response level, but after function_call_output the client must send response.create, no cancellation event is documented, and there is no filler setting. source |
| Pipecat B Daily | Per-tool cancel_on_interruption, timeout_secs, on_function_calls_started fillers and run_llm continuation exist, but the default cancels on any speech; an open issue shows a refund that landed reported as cancelled and some models refunding twice. source |
| Retell AI API + MCP B Retell AI Inc. | Speak During and After Execution, timeout_ms and error results cover fillers and continuation, but Retell's own custom-LLM guide says a response discarded when the caller keeps talking is re-requested, so a tool runs twice unless the developer adds an idempotency key. source |
| Rime Prism B Rime Labs, Inc. | prism_filler sets the phrase during a tool and failures are spoken from the result, but the backend must ask Prism to continue after each result, a tool cannot be forced, and Prism does not deduplicate requests. source |
| Synthflow API + MCP D AgentFlow AI GmbH | Custom actions have start, delay (threshold 0 to 30 s) and failure messages and a timeout up to 30 s, but the docs read say nothing on cancellation by caller speech, duplicate execution or forcing an action. source |
| Vaani Developer Platform E Vaani AI Research Private Limited (Bengaluru, India, CIN U62099KA2025PTC201212) | Tools are defined from a cURL request and triggered by a natural-language description, with configurable filler words, but nothing documents timeouts, duplicate execution, interruption or what the agent says on failure. source |
| Vapi API + MCP B Vapi Inc. | Filler messages (request-start, request-response-delayed after N ms) and automatic resumption after the result are documented, but no cancellation event, no toolChoice field, and timeouts are not retried while the server may still complete the write. source |
| Vocily AI C VOCILY AI PRIVATE LIMITED | Each action has a short message spoken before it runs and a message for failure, but no timeout, cancellation-on-speech or duplicate-execution behaviour is documented in the tools page or the OpenAPI file. source |
| Vogent API E Monoid, Inc. | FunctionLifecycleMessages give started and waiting lines, and awaitSpeech holds a function until the agent stops talking. Nothing documents a timeout, what happens when the caller interrupts a running function, or duplicate execution. source |
Couldn't check 2
The docs were closed to us, or silent. These carry no mark either way.
| Listing | Why |
|---|---|
| Harmony.ai E monday.com Ltd. | No page read documents how the agent runs actions mid-call, fillers, timeouts or cancellation; agent behaviour is set in the web app only. source |
| Hume EVI (Empathic Voice Interface) C Hume AI, Inc. | EVI docs withdrawn: every EVI page, including the overview and tool-use guide, redirects to the Audio Decisions overview, and the EVI AsyncAPI file returns 404. The last assessment recorded EVI access ending 13 November 2026. source |
Doesn't do this job 4
Listings in the category that do something else, with the reason.
| Listing | Why |
|---|---|
| CueLight D sshah03 (individual GitHub account, no name or organisation stated) | No tool calling or conversation; it returns an end-of-turn probability per 20 ms of audio. source |
| fone.me F Fone | A hosted personal agent with no developer-registered tools or webhooks; outside agents reach it only through message_agent and A2A. source |
| Pipecat PhoneLLM B Daily | The card covers tool calling through vLLM's qwen3_coder parser, but cancellation, fillers, timeouts and precisely-once execution belong to the voice framework, not the model. source |
| PlaceCall C VOYGR Tech, Inc. | The developer cannot register tools for the call agent; the only mid-call hook is ask_user, a question routed to the caller's agent. source |
What stops agents finishing
- default cancel-on-interruption with no per-tool override
- no cancellation event when a blocking tool is interrupted, so the model re-issues the call
- tool result held behind the model's own speech until the caller speaks again
Around it
Before: voice-agents.inbound-receptionist. After: booking.reservations, payments.refund, crm.update-record.
Where this comes from
Weight 18 of 100 in Conversational voice agents: 9 of 31 accounts are tool-call failures in flight (acc-02, acc-03, acc-05, acc-14, acc-16, acc-17, acc-22, acc-24, acc-26); the largest failure cluster found
The same as data: /api/v1/use-cases/voice-agents.tool-call-mid-call.json, or the MCP tool get_use_case with id voice-agents.tool-call-mid-call. Other jobs in this category: /api/v1/niches?category=voice-agents.
For companies
Do agents find, use and choose your tools?
An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.