AudioPod for Agents

We've spent the past few months making AudioPod work better for one specific user: the AI agent acting on someone's behalf.

That user can't click around a UI. It can't watch onboarding videos. It needs three things from any audio service: discoverability, a stable wire format, and one auth header that already works.

Today we're shipping all three.

What's new

1. A real MCP server at https://mcp.audiopod.ai (Streamable HTTP, MCP 2025-06-18). Nine audio tools — music generation, TTS, voice cloning, stem separation, transcription, denoising, translation, karaoke, format conversion — exposed to any client that speaks Model Context Protocol.
2. A bundled CLI in both SDKs. pip install audiopod or npm install -g audiopod and you can audiopod music "lo-fi rainy 90 BPM" --out song.wav from any terminal.
3. A single /for-agents landing page that documents every discovery surface AudioPod publishes — agent card, MCP server card, agent skills index, OAuth Protected Resource metadata, A2A profile, llms.txt, llms-full.txt, ai.txt.

Why this matters

Agents are how new audio tools will get discovered.

Look at what's actually happening in 2026: a writer asks Claude Desktop to "make a podcast intro that fits this script," and Claude consults its connected MCP servers to figure out who can do that. A creator in a Cursor session asks for "a 60-second uplifting bed for the YouTube intro" and the editor calls a tool. A community manager pastes 90 minutes of meeting audio into a CLI script and gets a clean transcript back five minutes later — without a single click.

The agents already know how to do this. They just need somewhere to call.

Plug into Claude Desktop in 30 seconds

1234567891011

{
  "mcpServers": {
    "audiopod": {
      "url": "https://mcp.audiopod.ai",
      "headers": {
        "X-API-Key": "ap_YOUR_KEY"
      }
    }
  }
}

Drop that into ~/Library/Application Support/Claude/claude_desktop_config.json (or the equivalent on Windows), restart Claude, and ask:

"Generate a 30-second lo-fi instrumental for studying. Save it to my Desktop."

Claude will list AudioPod's nine tools, pick generate_music, and stream you back a job id. Same setup works for Cursor, Continue, Cline, and Goose — the docs have copy-paste snippets for each.

From the terminal

If your workflow is more shell-shaped than chat-shaped, the CLI is the same nine tools through a different door:

123456789101112131415161718

# Authenticate once
audiopod login

# Generate music
audiopod music "lo-fi rainy 90 BPM" --duration 60 --out song.wav

# Voiceover
audiopod tts "Welcome to AudioPod." --voice 368 --out hello.mp3

# Transcribe a meeting with speaker labels
audiopod transcribe meeting.mp3 --diarize --format srt > meeting.srt

# Split a song into stems
audiopod stems track.wav --mode six

# Clone a voice from a 30-second sample
audiopod clone reference.wav --name "Narrator"

pip install audiopod ships it on PyPI; npm install -g audiopod ships it on npm. Same commands, same flags, same JSON output (--json) for piping into other tools.

Open standards everywhere

The whole point of being agent-native is that nothing is bespoke. Every discovery surface an agent might check is published, machine-readable, and documented:

• /.well-known/mcp/server-card.json — MCP server descriptor
• /.well-known/oauth-protected-resource — RFC 9728 OAuth metadata
• /.well-known/oauth-authorization-server — RFC 8414 issuer metadata
• /.well-known/openid-configuration — OpenID Connect Discovery 1.0
• /.well-known/agent-card.json — A2A protocol agent card
• /.well-known/agent-skills/index.json — agentskills.io catalog
• /.well-known/api-catalog — RFC 9727 API directory
• /llms.txt and /llms-full.txt — llmstxt.org

Every URL above returns 200 with a stable JSON shape. No closed schemas, no proprietary handshakes.

Auth that already works

We accept either Authorization: Bearer <jwt> or X-API-Key: ap_*. Both work everywhere — REST, MCP, CLI, and the SDKs. The OAuth Protected Resource document advertises every supported scope, so a client that wants to programmatically discover what permissions to request can do so without parsing English.

API keys can be scoped to specific tools (music:generate, voice:synthesize, transcribe, stems:separate, voice:clone, audio:read, audio:write). Wildcard * works for full access. Per-key scope enforcement happens at the request boundary, not in the agent.

Three concrete demos

1. Claude Desktop generating a song from a sentence

"Hey Claude, please make a 60-second instrumental that would work as a YouTube intro for a science channel — something curious and uplifting. Save it to my Downloads folder."

Claude reads the connected audiopod MCP server's tool list, picks generate_music, sends the prompt, polls for the result, and downloads the file. The user did one sentence of work.

2. Cursor scripting a multi-language voiceover pipeline

12345678910

import AudioPod from "audiopod";
const client = new AudioPod();

for (const lang of ["en", "es", "fr", "ja"]) {
  const out = await client.voice
    .speak({ text: greetings[lang], voice: 368, language: lang })
    .then((j) => client.voice.waitForCompletion(j.id));
  await out.download(`hello-${lang}.mp3`);
}

The agent inside Cursor wrote that, ran it, watched four files appear in the workspace.

3. Terminal one-liner

12

audiopod music "uplifting cinematic strings, 3 minutes, 90 BPM" --out theme.wav

That's it. No project, no configuration, no SDK setup. The CLI handles auth, dispatch, polling, and download.

Get started

1. Get an API key. Free tier credits unlock the full tool surface — no card. → Dashboard → API Keys
2. Pick your surface. MCP, CLI, Python SDK, or Node SDK.
3. Build something. The full developer landing page lives at audiopod.ai/for-agents — start there if you want one place that documents every option.

What's next

A few things on deck for the next two months:

• More tools. Audiobook generation, podcast generation from documents, and lyric-video creation are all already in the REST API; we'll wire them into the MCP server next.
• Server-Sent progress events. Long jobs (music, stems) currently return a job id; we're going to stream progress updates directly through the MCP transport so agents can show "75% complete" cards.
• Webhooks for job completion. So you can fire off a generation request, walk away, and have the result pushed to your endpoint when it's ready.

If you build something interesting with AudioPod for Agents, tell us. We're building this for you.