PyAI Voice AI
Transcribe recordings, generate speech and build interruptible website voice assistants with your coding agent.
Server MCP Terhosting
npx add-mcp 'https://api.pyai.com/mcp'Terpasang ke Claude Code, Codex, Cursor, dan lainnya
Dokumentasi
Natural voice. Now in your flow.
Make something you can hear. Build an agent that can listen. Bring PyAI into ChatGPT, Claude, Cursor, Codex and Devin with one connection.
One URL. Browser sign-in. Your chosen project.
Your app. Your PyAI project.
Browser sign-in
Connect Cursor to PyAI
For a workspace login, add the entry below to your project. In Customize → MCPs, select the workspace, enable its PyAI entry and choose Authenticate under Local. Select your project in the browser. Before any billable action, call whoami on that connection and check the project ID, environment and billing against your approved budget.
{
"mcpServers": {
"pyai": {
"url": "https://api.pyai.com/mcp?auth=required"
}
}
}
Sign in → select a project → approve access. Cursor setup reference
01 / Ask. Create. Keep going.
Eight ways to put voice to work.
Start with the result you want. Your AI app discovers the tools, checks the inputs and works with your authorized PyAI project.
Speak
A voiceover, from a sentence.
“Turn this welcome message into warm, natural speech and give me the audio.”
list_voices → synthesize_speech
Hear
From recording to readable.
“Transcribe this interview and help me find the key moments.”
transcribe_audio · transcription jobs
Agents
Build your next voice agent.
“Create a booking assistant with a greeting, clear boundaries and a human handoff.”
create_agent · tools · knowledge
Trace
Find the evidence in your calls.
“Find calls with missing disclosures and show the supporting evidence.”
interactions · findings · rule packs
Recap
A useful summary. Next steps, too.
“Summarize this sales call and extract objections and follow-up actions.”
create_recap → get_recap
Dub
Let your recording travel.
“Dub this English recording into Hindi and give me the finished audio.”
capabilities → job → audio
Cast
Give your script some direction.
“Turn this script into expressive narration. Preview a line before the full render.”
direct → preview → render
Calling
Preview the next call.
“Show me the number and call request without buying or dialing. Explain any permissions I still need.”
discover_tools → dry-run preview → human approval
03 / Give your agent a head start
Less guessing. More building.
Public Markdown, live tool discovery and the API contract give coding agents the details they need.
Use the connected PyAI MCP server.
Read get_started and check whoami.
Call discover_tools with query (or task) for my use case.
Inspect openapi_get before constructing API requests.
Preview purchases and dialing; obtain approval before submitting.
Use https://pyai.com/mcp-agent-guide.md for conventions.
Keep job IDs and return usable audio or transcripts.
Prefer local stdio? Reuse your CLI login.
npx -y @pyai/mcp@0.5.26 login
# Then configure your host to launch:
npx -y @pyai/mcp@0.5.26
Already ran pyai login? The MCP server uses that saved profile. Full local setup.
Give OpenClaw ears, a voice and a next step.
Install the skills you need, connect your PyAI account through the official CLI, and ask your assistant to put them to work.
Hear · v1.0.0
Turn a voice note into useful text.
Transcribe a supplied recording, interview or voicemail and return the words you can search, quote or work with.
A PyAI account and transcription permissions. Audio processing consumes usage.
Speak · v1.0.0
Give your next answer a voice.
Turn your text into a spoken update, an accessible audio version or a voiceover with a supported PyAI voice.
A PyAI account and speech-generation permission. Speech generation consumes usage.
Recap · v1.0.1
Keep the context. Find the next step.
Retrieve an existing recap or summarize a supplied customer/agent transcript with its original speaker roles and timing.
Recap enabled and recap:read access. New transcript processing needs suitable timed utterances; a plain transcript may need preparation.
Calling · v1.0.0
Let your assistant make the call you approve.
Prepare and preview a phone call, confirm who it will reach and why, then place it only with your authorization. Read the returned status and available artifacts.
An eligible PyAI project, telephony:manage, an owned caller number, a configured agent and credits. Real calls can incur charges; sandbox access alone cannot dial.
A few things worth knowing.
What is PyAI MCP?
PyAI MCP connects AI apps to the PyAI voice APIs through typed tools. Use it to generate speech, transcribe audio, manage voice Agent profiles, review compliance, summarize calls, dub recordings and render directed narration.
Do I need to copy an API key?
The remote connection uses browser OAuth. Sign in to PyAI, select a project and approve the requested permissions. Your AI app receives a scoped connection; your PyAI API key stays on the server.
Can I connect in ten seconds?
An existing signed-in user can go straight to project consent. Setup time also depends on your app, account and workspace policy. A new account or an administrator approval can take longer.
Will it appear automatically in ChatGPT or Claude?
Use a custom app or connector to add PyAI by URL. Public app-directory listings require separate submission and review. Your host’s plan and workspace policies determine whether custom connections and write tools are available.
Can MCP run a live voice call?
MCP can manage Agent profiles and prepare or submit managed telephony requests. Buying a number or dialing requires telephony:manage, explicit confirmation and an idempotency key. Unconfirmed requests return needs_human without submission. Realtime microphone audio uses the SDK or widget; creating a profile does not place a call.
Can I test calling with a sandbox key?
Use dry-run previews to inspect requests without buying or dialing. Live number search and listing require telephony:manage, which a sandbox key may not include; a 403 means permission is missing. A preview does not prove a real phone call will connect.
How do updates work?
PyAI updates the hosted endpoint through its deployment process. Local @pyai/mcp installations have their own package releases and need restarting after an upgrade. The CLI ships in @pyai/sdk; installed versions do not silently upgrade when a new SDK is released.
How do I disconnect an app?
Open Connected AI apps in the PyAI console and revoke the connection. Access lasts up to 30 days and also stops when the authorizing user loses project access. Revocation stops new requests; already-submitted jobs may continue.
What does it cost?
The MCP connection uses your project’s normal PyAI credits, product pricing and limits. Generated audio, transcription, dubbing and other processing consume usage. Reading the in-server setup guide and workflow recipes does not consume model credits.