AI Voice Agent
August 26, 2026 · View on GitHub
An AI agent for voice, narration, dubbing, and calling — text-to-speech, voice cloning, multilingual dubbing, outbound calling, and podcast production — backed by real voice-generation APIs.
Part of Agency Agents OS, an open ecosystem of specialized AI agents for real business work.
Related Projects
- Agency Agents OS — the central catalog this repo is part of.
- ai-video-agent — pairs this repo's narration/voice output with lip-synced avatar video.
- ai-image-agent — generates the avatar reference images this repo's avatar-ugc sub-agent animates.
- MuAPI music and speech docs — the underlying voice/speech model catalog.
- MuAPI MCP docs — connect this repo's
SKILL.mdfiles via MCP. - MuAPI access keys — create the API key this agent needs.
What this covers
This repo is the umbrella for anything an agency or in-house team would call "the AI voice agent": producing narration, dubbed audio, or a cloned voice for content and outreach.
Sub-agents
| Agent | Does | Status |
|---|---|---|
| Narration / Dubbing | Script-to-speech narration and multilingual dubbing | Blueprint |
| Voice Cloning | Clone a voice from a reference sample for consistent narration | Blueprint |
| Outbound Calling | AI-voice outbound phone calls for outreach/qualification | Coming Soon |
| Podcast Strategy | Turn source material into a scripted podcast episode plan | Blueprint |
Required Muapi APIs
media.generate_speech— text-to-speech narration across Muapi's voice catalog.media.clone_voice— voice cloning from a reference sample.media.dub_audio— translate and re-voice existing audio.
See each sub-agent's SKILL.md for the specific capabilities it uses.
Setup
- Create a Muapi account and API key at muapi.ai.
- Review the Muapi API quickstart and OpenAPI schema for the voice-generation endpoints.
- Load the
SKILL.mdfor the sub-agent you need into your agent runtime (hosted agent, MCP client, or custom LLM app), or follow it manually.
Using with an AI agent
Every sub-agent's SKILL.md is model- and runtime-agnostic — it's plain Markdown, so it works with any LLM agent, not just Claude. Two integration paths:
As an MCP connection (the agent gets live Muapi tools):
Muapi runs an MCP server at https://api.muapi.ai/mcp that any MCP-compatible client can connect to — Cursor, Windsurf, Claude, or your own custom agent.
- Cursor / Windsurf / other clients with a header field: connect to
https://api.muapi.ai/mcpwith anAuthorization: Bearer YOUR_MUAPI_KEYheader. - claude.ai / Claude Cowork / other connector UIs with no header field: use the URL-embedded key form instead,
https://api.muapi.ai/mcp/YOUR_MUAPI_KEY, via Settings → Connectors → Add custom connector. - Claude Code / Claude Desktop:
claude mcp add muapi -e MUAPI_API_KEY=YOUR_MUAPI_KEY -- muapi mcp serve(uses the muapi CLI's stdio transport — Claude Code's HTTP MCP client doesn't reliably inject tools).
Full setup details for every client: muapi.ai/docs/mcp.
As agent instructions (any LLM follows the workflow directly):
Drop a sub-agent's SKILL.md into a Claude Code project's .claude/skills/ directory, paste it into a custom-GPT/Project's system instructions, hand it to an autonomous agent framework as a tool spec, or attach it directly in a chat conversation — then ask the agent to follow it.
Read-only vs. write actions
Narration, dubbing, and voice cloning are draft-only — output is an audio file, not a publish. Outbound calling is requires-approval-to-publish since it contacts a real person; see that sub-agent's approval boundaries.
Status and limitations
Narration and dubbing are live, tested Muapi capabilities. Voice cloning requires explicit consent from the voice's owner — see the sub-agent's approval boundaries. Outbound calling (Concept) depends on telephony/calling API coverage not yet live on Muapi.
Contributing
See Agency Agents OS CONTRIBUTING.md.