Podcast Track Cleaner Pro can be driven by an AI agent over the Model Context Protocol. You connect Claude Code, Claude Desktop, Cursor or any other MCP client to the app, ask it to clean the open episode, and watch it work: every edit happens in the project window in front of you, and every one is an ordinary undo step.
Agent control is in Podcast Track Cleaner Pro, the edition distributed directly. The Mac App Store edition is sandboxed and leaves it out.
Turn it on
Allow agents
Open Settings ▸ Agents (⌘,) and tick Allow agents to control Podcast Track Cleaner Pro. The status line should read Listening on 127.0.0.1:7419.
Note the token
The app makes a random token; an agent has to send it with every request. New Token replaces it and cuts off every agent set up with the old one.
Connect your agent
Use one of the two copy buttons on the same page — the token is already filled in — and paste it where the next section says.
The server only listens on this Mac (127.0.0.1), refuses requests without the token, and refuses requests a web page tries to make from a browser. If port 7419 is taken by something else, the status line says so; type another port and press Apply, then copy the setup again.
Connect an agent
Claude Code
Click Copy Setup Command and run it in Terminal. It looks like this:
claude mcp add --transport http track-cleaner http://127.0.0.1:7419/mcp --header "Authorization: Bearer YOUR_TOKEN"Start claude in any folder and /mcp should list track-cleaner as connected. Add --scope user to the command to have it in every project.
Cursor, LM Studio and other apps
Click Copy JSON Configuration and paste it into the app's MCP settings (for Cursor, ~/.cursor/mcp.json):
{
"mcpServers": {
"track-cleaner": {
"type": "http",
"url": "http://127.0.0.1:7419/mcp",
"headers": { "Authorization": "Bearer YOUR_TOKEN" }
}
}
}
Claude Desktop
Claude Desktop starts its local servers as commands, so it reaches the app through the mcp-remote bridge (it needs Node.js). In Claude Desktop, Settings ▸ Developer ▸ Edit Config, add this, and restart it:
{
"mcpServers": {
"track-cleaner": {
"command": "npx",
"args": [
"-y", "mcp-remote", "http://127.0.0.1:7419/mcp",
"--header", "Authorization: Bearer YOUR_TOKEN"
]
}
}
}
Any model will do that your client can run with tools, but a cleanup is a long job of dozens of careful steps, so a strong model makes a real difference. Small local models tend to lose the thread halfway through.
Ask for a cleanup
Open the episode in the app first — the agent always works on the front project window. Then ask in plain words:
Clean up the open episode: remove mouth clicks and loud breaths, even out the two speakers, and tell me what you changed.
For the whole house-style cleanup, the app also offers a ready prompt, clean_episode. In Claude Code it is the slash command /mcp__track-cleaner__clean_episode; other clients list it with their prompts. It tells the agent to read the app's cleanup guide first, follow it step by step, show a dry run before every bulk edit, and finish with a report — and not to save or export unless you ask. Add notes for it after the command: where the show really starts, what to keep, where a transcript is.
A good session goes roughly like this:
It reads the guide and the project
The agent learns the order of work — trim, align, silence, un-overlap, level, parasite words, assets, export — and the targets: pause lengths, levels, how much of the show a cleanup usually removes.
It measures instead of listening
Agents can't hear. They read each track as numbers — loudness over time, the noise floor, the speech level, the quiet stretches and the loud peaks — and decide from those.
It edits, one undo step at a time
Every tool calls the same code as the menus, so the edits appear on screen as they happen and land in the undo history. The two jobs that take hundreds of hand edits — tightening pauses and levelling each speaker's turns — are single steps.
It reports back
What changed, before-and-after numbers, what it skipped, and timestamps worth listening to. It can select a region and play it for you to show what it means.
What the agent can do
| Tools | What they do |
|---|---|
get_project, analyze_track | Read the tracks, selection, chapters and undo state; measure levels, noise floor, pauses and peaks |
get_cleanup_guide | The house cleanup style, measured from episodes cleaned by hand |
find_marks, list_marks, set_marks, apply_marks | Find clicks, breaths, throat clears and quiet noise; veto the doubtful ones; apply the rest |
edit_region | Silence, delete, gain, normalize, fade, level-line, noise reduction, beep, remove overlap, split |
process | Auto-level, tame peaks, silence background, auto-align, trim silence from the ends |
tighten_pauses, normalize_turns | Shorten long pauses on all tracks at once; bring each speaker's turns to the target level |
edit_track, chapters, assets | Rename, mute, move, merge tracks; add and name chapters; place intros, outros and beeps |
select, playback, history | Show you a region, play it, undo or redo |
open_project, add_tracks, save_project, export_mix | Files — only when you ask for them |
Before you publish
Listen to it
Agents judge audio by its levels. A cut that looks clean in the numbers can still land mid-breath, and a quiet "yes" can look like noise. Play the timestamps it gives you, and skim the rest.
Undo is yours too
Agent edits and your own share one history. ⌘Z takes back the last thing either of you did, and you can keep editing by hand while the agent is connected.
Save before you start
Undo goes back twenty-four steps by default (Settings ▸ General can raise it), and a cleanup is often more. Save the project before handing it over (⌘S); if you don't like where it ended up, close without saving and you are back.