How to edit a screen recording with an AI agent

What you need

• Cadre on macOS 13 or later, with a project you've recorded.
• Claude Code, or any MCP-capable agent client.
• A minute to run one claude mcp add command.

That's it. Cadre's editor is fully data-driven — a project is a manifest plus timeline data — so the app can expose the whole edit as tools an agent calls. You describe the outcome; the agent chooses the tools and verifies each change.

1. Record your demo

Open Cadre and record as normal: your screen with system audio, mic and webcam, or an iPhone/iPad mirrored over USB. As you record, Cadre tracks every click, keystroke and cursor movement. That interaction data is what makes the next part possible — the agent (and Cadre's own auto-zoom) can see where the action was, not just the pixels.

Recording stays something you do by hand, on purpose. You know what you want to show; the agent takes over once the take is done.

2. Connect the agent

When Cadre is running it writes a connection file to your app-support directory with the current port and a fresh bearer token, then removes it on quit:

{
  "port": 51837,
  "token": "a7Q2…base64url…",
  "version": "1.0"
}

Point Claude Code at it with a single command:

$ claude mcp add --transport http cadre \
    "http://127.0.0.1:$PORT/mcp" \
    --header "Authorization: Bearer $TOKEN"

The server binds to 127.0.0.1 only, never a public interface, and checks the bearer token on every request. Once it's connected, you talk to your agent in plain language — no need to know the tool names. Full connection details, including a project-local .mcp.json, are in the agent docs.

3. Direct the edit

Now describe what you want the way you'd brief an editor. Under the hood the agent selects from Cadre's v1 tools — here's how a few common requests map to real calls:

• It reads the current state with get_timeline (durations, zooms, cuts, speeds, captions).
• It removes dead air with add_cut and ramps slow stretches with set_speed.
• It places cinematic zooms with add_zoom — Cadre's spring physics handles the push-in and ease-out.
• It transcribes on-device with generate_captions and styles them with set_style.
• It renders with export_video and polls get_export_status until the MP4 is ready.

Every tool validates its arguments and returns the post-edit entity, so the agent can confirm each change landed the way you asked — and undo it with undo if you don't like it.

Prompts that actually work

You don't call tools; you describe outcomes. These map cleanly onto the v1 surface:

What a full session looks like

One prompt can carry the whole edit. Here's the shape of it — you say one sentence, the agent reports back what it did:

# you just finished recording a 4-minute product demo
› open my latest project, tighten it, zoom the terminal
  whenever I type, caption it, and export at 1080p

# the agent works through the tools and reports back
● open_project      loaded "api-walkthrough"
● get_timeline      4:12 · 0 zooms · 0 cuts
● add_cut           trimmed 3 silent gaps  −0:47
● set_speed         2× ramp over a 22s idle stretch
● add_zoom          6 zoom keyframes on typing bursts
● generate_captions 38 segments · on-device
● export_video      rendering… get_export_status 100%
✓ demo.mp4  1080p · H.264 · 3:25

You reviewed a four-minute take and shipped a tight, captioned, cinematically zoomed cut without opening the timeline yourself. If you want to see exactly which tools exist and what each returns, the tool reference lists the full v1 set.

When to still edit by hand

Agent editing isn't all-or-nothing. The agent is great at the mechanical passes — cutting dead air, placing zooms on activity, captioning, exporting. But taste is yours. Open Cadre's editor directly when you want to:

• Nudge a single zoom's framing or timing to land a beat exactly right.
• Fine-tune the background, device frame or webcam position for a specific shot.
• Scrub and preview before committing to an export.

A good flow is agent-first, hand-finished: let the agent do the heavy lifting in one sentence, then polish the two or three moments that carry the demo. Everything the agent does is a normal, reversible edit on the same timeline you'd touch by hand.

FAQ