How to Edit Instagram Reels With Claude AI (Edits Desktop MCP Setup)

On November 5, 2026, Instagram shipped Edits for Mac. Almost every write-up covered the same things: free, 4K, no watermark, CapCut had better watch out. One line in the announcement got almost no coverage, and it is the most interesting thing in the release:

“And for anyone using AI agents, you can connect them to the desktop editor through MCP so they can help with your projects.”

That is Instagram saying, in a product blog, that an AI agent can drive its video editor. Not generate a video. Edit one — place clips, land cuts, stack overlays, style text, trigger the export. This guide is how to actually set it up, what the server can and cannot do, and where it falls over.

What you need before you start

Three requirements, all hard:

• A Mac with Apple silicon (M1 or later). Edits on desktop is macOS-only. No Windows, no Linux, not even Intel Macs. If you're on a PC, this is the situation.
• Edits for desktop, free from Instagram, signed in with your Instagram account.
• Claude Code running on the same machine. The automation server listens on loopback only. Claude in a browser tab cannot reach it; neither can a cloud agent. This has to be local.

Turn on the automation server

Inside Edits, open Settings → Automation and switch on the automation server. The app starts listening on http://127.0.0.1:19501/mcp and shows you a bearer token.

You can confirm it's live from a terminal:

lsof -nP -iTCP:19501 -sTCP:LISTEN

Edits  91109  TCP 127.0.0.1:19501 (LISTEN)

Note the address. 127.0.0.1 means loopback — nothing outside your Mac can connect, and nothing is being sent to Meta by the server itself. That matters for the risk conversation later.

Connect Claude Code in one command

The Automation settings page gives you the command. It looks like this:

claude mcp add --transport http --scope user edits \
  http://127.0.0.1:19501/mcp \
  --header "Authorization: Bearer <your-token>"

Run it, start a new Claude Code session, and the editor tools load as mcp__edits__*. That's the whole setup. No JSON config to hand-edit, no Python environment, no scraping library — which makes this the least painful MCP install in the creator space by a distance.

What the tools actually cover

Meta publishes no documentation for this server, so the only honest way to describe it is to ask it. Querying a running instance for its capabilities on November 13, 2026 (schema 2026.10.19-1) returns the catalogs the agent can draw from:

The server also tells an agent how to think about the timeline: time in seconds, canvas coordinates normalised 0–1 with the origin top-left, rotation in degrees clockwise, colours as #RRGGBB. It even ships worked recipes — an animated title is addText then applyTextPreset then setTextAnimation; a beat-synced cut is beatMarkers then splitClip.

That last one is the tell. Beat markers plus clip splitting is how you get an agent to cut on the music instead of on a guess.

What a realistic session looks like

The workflow that people are reporting, and the one that matches what the tools are good at:

1. Stage your media. Edits only reads from its own sandbox directory, so clips and overlay PNGs get copied there first.
2. Describe the video, not the timeline. What it's about, how long, what the beats are. The agent picks the cuts.
3. Let it assemble. Overlays at timestamps, sound effects on the beat, a styled title, auto captions in one caption style for the whole video.
4. Have it check its own work. The server can return preview frames, so the agent can look at the frame it just built before you do.
5. You finish it. Open the project, fix the hook, watch it start to finish, export.

Nate Curtiss, who runs a YouTube agency, put the division of labour better than anyone: “keep 2 parts for yourself: the hook — the first 5 seconds decide whether anyone watches the rest. the final review — watch it start to finish before it goes out.”

He's right, and the reason is structural. An agent optimises the assembly, which is the boring, mechanical, hour-of-clicking part. It has no idea whether your first five seconds earn attention, because that is a judgement about your audience that lives in your head.

The five limits that will bite you

It cannot post, and that is the point

The automation server does not publish. No posting, no trial reels, no scheduling. You export and share from the Edits UI yourself.

People read that as a missing feature. It is the best thing about it. We've written before about how most Instagram MCPs are a ban risk wearing a product costume — the viral ones run on unofficial scraping libraries and ask for your raw Instagram password, which is the single most reliable way to lose an account.

Where the automation actually ends

Worth being clear-eyed about what you have and haven't automated here. Editing: yes, the assembly half of it. Publishing: no, by design. And the part that decides whether the reel earns anything — replying to everyone who comments asking for the link — is not something a video editor can touch at all.

That gap gets wider the better this works. Ship three times the reels and you get three times the comments. The follow-through is its own system, built on Instagram's official API, and it is the half that most people skip.

Should you bother?

Honest answer: it depends entirely on whether assembly is your bottleneck.

If you film your own content, post regularly, and lose an hour per reel to clicking overlays into place, this removes most of that hour. If you already work with an editor, or you only post occasionally, the setup cost is not worth it yet — the tooling is eight days old, undocumented, and changes weekly.

What's not in doubt is the direction. An Instagram product now ships an agent interface to its own timeline. That is a fairly loud signal about where the editing stack is heading, and it is worth an afternoon to understand before everyone else does. For the parts you still do by hand, our Instagram Edits help guides cover the manual workflows.