Podcast Editing AI That Asks Before It Spends: Let Claude, Cursor or n8n Cut Your Episode Through Cast

Podcast editing AI in Cast plays you 30 seconds of your own episode, already edited, about a minute after upload. Now Claude, Cursor or n8n can run it for you through one URL, https://cast-mcp.mubert.com/mcp, approved once in the browser. That’s it. The agent then cuts the episode to the length you asked for, puts music under it and exports. This post covers podcast editing with AI from setup to export for people who cut their own shows: what the agent can do, what it can’t, and the switch that keeps it away from your credits.

Editing a talk show by hand is forensic work. You zoom the waveform until one “um” looks like a mountain range. You delete it, play it back and find the guest now sounds like a time machine interrupted them. We did this for years. Our worst evening went on one sentence, which we then cut entirely.

Podcast editing AI now plugs in through MCP, not a plugin

Cast speaks MCP, the Model Context Protocol, which its own docs call “an open-source standard for connecting AI applications to external systems”. In plain words, it lets an assistant press buttons in another app instead of telling you which buttons to press. Cast’s docs say any MCP client works: Claude Code, Claude Desktop, Cursor, Windsurf, VS Code, n8n, or your own code.

The MCP docs compare the protocol to a USB-C port. That is fair, although USB-C has never asked us whether it may spend money.

For podcast editing, AI used to mean a magic button inside one app, and you clicked it there or nowhere. With MCP, the podcast editing itself becomes a set of tools any AI assistant can call: import, transcribe, plan cuts, build the project, export. So the assistant that already writes your show notes can also cut the show. Cast’s docs even list “many episodes at once” among the jobs an agent handles. A whole season, then.

There is a catch, and Cast’s docs state it themselves: “an agent spends faster than a person”. We come back to that below, because trust in the whole setup depends on it.

Connecting Claude, Cursor or n8n to Cast takes one URL

Setting up podcast editing AI in Cast takes one address and one approval in the browser. No token to copy, unless your client can’t open a browser. The first time the assistant uses the server, a Cast page opens. Sign in and click Allow. Cast then creates a token for that one assistant and lists it in Settings → API & MCP, where you can revoke it any time.

For Claude Code it is one command:

claude mcp add --transport http cast https://cast-mcp.mubert.com/mcp

The other clients differ in where the URL goes:

• Claude Desktop and claude.ai: Settings → Connectors → Add custom connector, paste the URL, click Connect.
• Cursor, Windsurf and VS Code: add a cast entry with "type": "http" and the URL to the mcpServers block of the client config (Windsurf calls the field serverUrl), restart, approve.
• n8n, CI jobs and scripts: create a token in Settings → API & MCP and send it as an Authorization: Bearer <token> header to the same endpoint.

For the third case, n8n’s MCP Client Tool node lists Bearer among its authentication options. The node also lets you expose all of a server’s tools, only selected ones, or all except some. That is handy for an n8n podcast editing flow where the AI may only transcribe and export. If you don’t want MCP at all, Cast’s docs say the API underneath is plain REST and the MCP server is “a thin layer over it”.

One more step is worth ten seconds. The tools carry the rules, but the cast-mcp skill carries the strategy: the order of calls that keeps a session cheap, and asking you before anything is paid. The install command is on the connect page. Skip it and the assistant still works, only less politely.

What the agent does to an episode, step by step

You ask the AI for one outcome, and it runs about seven podcast editing steps on its own. The example in Cast’s agent docs is: “Make a 15-minute episode from this 33-minute interview, keep the intro, put a calm bed under it.” (A bed is quiet background music that sits under speech.)

From that one sentence, the agent imports the file and transcribes it. It plays you the First Pass preview. If you say yes, it applies First Pass to the whole episode. Then it cuts the result down to your 15 minutes, builds the project and exports it. Every answer it gives you ends with a link to the project, so you never have to wonder where your episode went.

Cast’s docs split the podcast editing work into three layers, and AI handles only the first two:

• First Pass, Cast’s own editor, listens to the audio and does the cleanup.
• The agent follows your instructions: length, topics to drop, parts to keep, many episodes at once.
• The Cast editor is where you finish with full control: swap or regenerate the music, set the ducking, fix any word as text.

Ducking means the music automatically drops in volume whenever someone talks. Cast does it by default.

So the agent is not the editor here. It is the producer who holds the clipboard, books the session and makes sure the 33-minute interview comes out at 15. The listening happens elsewhere, which leads to the most useful fact in the whole setup.

First Pass hears the recording, the agent only reads it

An agent without First Pass works from the transcript alone. It can’t hear the recording, so it cuts only clear hesitations and leaves “so”, “like” and “well” alone unless you name them. That is the biggest quality gap in podcast editing AI: hearing versus reading. Cast’s docs put it plainly: “For the cleanest result, let it use First Pass.”

A transcript is a lossy copy of a conversation. “Like” can be a filler or a verb. A pause can be dead air or the beat before a punchline. First Pass listens to the audio, so it catches every um, false start and repeat, and keeps the words that carry meaning. It also writes chapters and picks a music bed for the kind of speech it heard.

The details are sensible. First Pass shortens long pauses to a natural beat rather than removing them, so the conversation keeps its rhythm. It takes out fillers and false starts where they interrupt the sentence, then cleans the voice and levels it to broadcast loudness.

A short tangent. “So” may be the most defended word in podcasting. Some hosts open every answer with it and would rather lose a guest than lose the “so”. An agent working from text cannot tell which kind of host you are. Anyway.

First Pass is free once, then priced per started 10 minutes

Your first pass on a new account is free, and the free plan gets one pass a month. After that, a preview runs only when you press Run, with its price on the button. Applying First Pass to the whole episode needs Lite or higher and is priced per started 10-minute block. A block counts once the episode is a minute into it.

We state the rules here, not the amounts. Check the live pricing page for credits, because we won’t print a number that might change next week.

Every cut is a proposal until you say build

Nothing destructive happens first, which is the rule that makes podcast editing by AI bearable. Cut planning is a dry run. You see every proposed cut with its timestamp and the exact words being removed, and nothing touches the audio until you say build.

You steer it in plain language. The docs’ own example: “Remove every ‘you know’ and shorten pauses over a second, but do not touch the first ninety seconds.” That is a sentence you would say to a human editor, and you get the same kind of result. The difference is that this editor shows its work before it does it.

When the build lands, it lands as ordinary work. The cuts appear on the timeline and also as word-anchored edits in the Voice Editor’s transcript. Word-anchored means each cut is tied to the words it removed, not baked into a re-rendered file. So you can reopen the episode and keep editing it as text, as if you had made every cut yourself.

Undo is generous too. Applying First Pass to the whole episode counts as one undo, and Cast saves a checkpoint before it. The Pauses and Fillers panels still list every single cut, so you can restore them one by one. We once kept a folder of files called things like final_v3_REAL_thisone.wav for exactly this purpose. A checkpoint is a calmer way to live.

An AI podcast editing agent cannot spend credits unless you tick a box

By default, an AI agent editing your podcast can’t spend anything. When you connect an assistant, the Cast page shows a “Let it spend credits” box that is unticked. Tokens you create yourself have the same switch. Even when you tick it, every paid step quotes its price first and spends only when the agent confirms that exact quote.

A read-and-edit connection can still do everything that’s free in the app. It can transcribe, analyse, plan cuts, enhance the voice, pick music you already own, build the project and export within your plan. It can’t buy anything. If it tries, the call comes back as a refusal with the price attached, so the agent can tell you what the step would cost.

A spending token adds the paid parts: First Pass previews, whole-episode edits and generated full-length music. These cost the same as in the app.

We like this part most: the check doesn’t rely on the assistant’s manners, because the Cast MCP server enforces it. Its first call returns a price and a quote id, and nothing is charged until a second call sends that id back. Think of a teenager borrowing your card who has to read you the receipt before every purchase, and whose card stops working the moment you say so.

That last part is literal. Revoke a token in Settings → API & MCP and it stops working at once, so the assistant has to connect and be approved again. Tokens also expire after a year either way, and a run that fails is refunded.

The music bed comes licensed and stays inside Cast

Music is the riskiest part of podcast editing with AI, so start with the licence: the bed is licensed music, generated on the same platform. An agent can reuse a track you already generated, search staff picks by describing a vibe, or render a new full-length bed. Cast then ducks it under speech on its own. On Plus and Max, exports include the creator commercial-license PDF, the document you keep in case anyone later asks where the music came from.

This matters more than it sounds. Most podcast music trouble starts with a track someone called “free” without reading the terms. We wrote a whole post about what “no copyright music” really means and where it comes from, and the short version is: keep the paperwork. The samples behind Cast’s music come from real artists, who are paid for them. So the bed under your interview has a supply chain, and the PDF describes it.

One honest limit. An agent can’t hand you the bare music file. The bed goes under your episode, and the track itself stays inside Cast. If you hoped to reuse your podcast bed as a YouTube intro, this pipeline is not for that.

Where podcast editing with AI still needs your ears

The agent does the boring first pass, and you make the final call by ear. That isn’t a disclaimer we tacked on. It is the design: Cast’s docs describe the editor as the place where you “finish by ear”. AI for podcast editing works best when you treat it like a fast assistant whose work you check, not like a replacement for listening.

Finishing covers the rest. You add chapters and set the loudness for the platform you publish to. Export formats depend on the plan: MP3 on every plan, WAV from Plus, per-lane stems (a separate file for each track lane) on Max. Choose before you plan a workflow around stems you don’t have.

Podcast editing AI fails in predictable places, so here they are without gloss. The agent works faster than you can check it, so actually read the dry run instead of skimming it. Without First Pass, fillers stay. On a long episode, most of the wait is the transcription behind it; on a short take it is under a minute. And a batch of ten episodes means ten dry runs to look at. Still faster than ten evenings with a zoomed waveform, but not nothing.

There is also a quieter benefit. Listening, comparing and editing by hand stay free on every plan, so the agent is an option, not a toll gate. You can let it run First Pass, take the project and do everything else yourself. Or do the reverse: cut by hand and ask the agent only for chapters and the export.

A checklist for AI podcast editing this week

1. Add the Cast MCP URL to your assistant and approve it in the browser.
2. Leave “Let it spend credits” unticked for the first session.
3. Install the cast-mcp skill before you ask for anything.
4. Ask for an outcome in one sentence: target length, what to keep, what kind of bed.
5. Spend your free first pass on a real episode and A/B the preview against the raw take.
6. Read every cut in the dry run before you say build.
7. Open the project in the Cast editor and fix anything that sounds off by ear.
8. Revoke the token in Settings → API & MCP if you stop using the assistant.

Podcast editing with AI is most useful on the one sentence you would have lost an evening to. Let the agent propose the cut, read it, and keep the “so” if you are that kind of host.