Top 10 Voice AI & Conversational Agent MCP Servers (2026)

create_call(assistantId?, phoneNumberId?, customer?: object, scheduledAt?)

Place an outbound call, optionally scheduled with an ISO datetime.

get_call(callId)

Fetch details of a specific call.

list_calls

List recent calls, limited to 10 per request in the handler.

create_assistant(name?, firstMessage?, model?: object, voice?: object)

Create a voice assistant with its model and voice configuration.

Limitations

• list_calls takes no arguments and the handler caps results at 10, so an agent cannot page through call history.
• Creating an assistant means supplying nested model and voice objects, which is hard for an agent to get right without reading Vapi's own docs first.
• Calls cost money per minute. An agent that retries a failed call is spending real budget with no guard in the server.

list_api_endpoints(search?)

List Retell API endpoints with name, HTTP method, path and a short summary.

get_api_endpoint_schema(endpoint)

Fetch the full JSON Schema for one endpoint's arguments.

invoke_api_endpoint(endpoint, arguments?: object)

Call an endpoint with arguments matching the schema returned above.

Limitations

• Any action takes three round trips: discover, fetch schema, invoke. Latency per action is higher than a server with direct tools.
• The agent has to pick the right endpoint from a list rather than from a typed tool, which moves the failure mode from a schema error to a wrong choice.
• Retell's docs describe capability categories and do not publish the endpoint list, so the only reliable inventory is the one the server returns.
• The Model Context Protocol registry lists a third-party Retell wrapper and not this server, so registry-led searches find the wrong one.

Limitations

• The endpoint requires an OAuth bearer token and returns nothing to an unauthenticated tools/list, so the tool inventory cannot be checked before you have an account.
• ElevenLabs documents capabilities in prose without publishing tool names or input parameters.
• The archived elevenlabs-mcp package still installs and still appears in most third-party setup guides. Check which one your config points at.
• Audio returned inline as base64 consumes context quickly, which is the recurring complaint across every speech server on this list.

text_to_speech(transcript, voice_id, output_format: object, model_id?, language?, speed?: number, volume?: number, emotion?, save?: boolean)

Synthesise speech with optional speed, volume and emotion control.

speech_to_text(file_path, mode?, model?, language?, sample_rate?: number)

Transcribe an audio file in batch mode by default, or streaming.

clone_voice(file_path, name, language, mode, description?)

Clone a voice from an audio sample.

localize_voice(voice_id, language)

Adapt an existing voice to another language or dialect.

Limitations

• Both speech_to_text and clone_voice take a file_path, so the server needs filesystem access to audio the agent cannot fetch for itself.
• The hosted endpoint at mcp.cartesia.ai rejects unauthenticated requests, so the tool list can only be read from the package.
• Admin API keys use a different prefix from standard keys and unlock different tools, which makes behaviour depend on which key you configured.
• Repository visibility is low at 14 stars, which understates a server that Cartesia maintains actively.

create-agent(body: object)

Create a voice agent.

deploy-a-branch(path_agentId, path_branchId, body: object)

Deploy a branch of an agent to an environment.

promote-a-deployment-to-the-next-environment(path_agentId, body: object)

Promote a deployment from one environment to the next.

trigger-an-outbound-call(path_agentId, body: object)

Start an outbound call from a deployed agent.

Limitations

• 106 tools is more than some clients accept. Cursor caps the tool list at 40, so the server needs filtering before it fits.
• Most tools take a generic body object, so the agent still has to know the request shape from PolyAI's API docs.
• Endpoints are region specific and the Builder and Data servers are separate, so a full setup means configuring two servers.
• Rollback and promotion tools change what is live for real callers, with no confirmation step in the server.

tts(utterances: array, voiceName?, provider?, continuationOf?, quiet?: boolean, modelVersion?)

Generate speech from utterances, each with optional delivery description, speed and trailing silence.

save_voice(generationId, name)

Save a generated voice under a name for reuse in later requests.

play_previous_audio(generationId)

Replay audio from an earlier generation by its id.

list_voices

List saved and provider voices.

Limitations

• Model version 1 supports delivery descriptions and generating without a voice. Version 2 supports multiple languages. Neither does both.
• Audio is written to a working directory on disk, so the server needs somewhere to write and the files accumulate.
• Five tools cover generation and voice storage only. There is no transcription and no telephony.
• Voices saved by name are the only reuse mechanism, so losing the generation id means regenerating the audio.

generateBXML(verbs: array, auto_gather?: boolean)

Build valid BXML from verb descriptions such as SpeakSentence, Gather, Transfer and Record.

respondToCallback(call_id, bxml)

Answer a live call callback with BXML.

listPhoneNumbers(size?: number, status?)

List phone numbers, defaulting to 100 results with status Inservice.

getCallbackEvents(call_id?)

Read callback events produced during a call.

Limitations

• There is no tool that places a call. The server builds and answers call control markup, so the call has to be started elsewhere.
• generateBXML takes a list of untyped verb dictionaries, so an incorrect verb shape fails at call time rather than at schema validation.
• Credentials are set through a tool at runtime, which means they pass through the conversation unless supplied as environment variables.
• BXML is specific to Bandwidth. The knowledge does not transfer to another carrier.

Limitations

• Bland publishes tool names without input parameters, and the endpoint rejects unauthenticated requests, so the schemas cannot be read before signing up.
• Outbound calling at volume runs into regulation that the server does not model, including consent and calling hours.
• Pricing is per minute, so a loop that redials is a loop that spends.

synthesize_speech(text?, speaker?, modelId?, language?, include_audio?: boolean)

Synthesise speech, with audio optionally excluded from the response.

list_voices(model?, language?, limit?: number)

List available voices, filtered by model and language.

generate_integration(framework, language?, model?, speaker?, mode?)

Generate integration code for a named framework using the chosen voice.

normalize_text(text)

Show how text will be normalised before synthesis.

Limitations

• Rime positions the server for exploration and prototyping rather than production traffic.
• Browsing voices works without a key, but synthesis needs one.
• The server is stateless, so nothing carries between calls and the agent resends context each time.
• generate_integration produces code for the frameworks Rime supports, so an unusual stack gets nothing useful.

Limitations

• Deepgram does not publish tool names or input parameters, describing the surface as every CLI command.
• The dg CLI has to be installed and authenticated first, which is a step the other entries on this list do not require.
• Tool behaviour follows CLI releases, so an upgrade can change the surface without a server version to pin.