Agentic video editing — edit by chatting

Editing without dragging clips

Traditional editors make you do the work: scrub the timeline, trim each clip, nudge keyframes, hunt for the right transition. Agentic editing flips that around. You say what you want — "shorten the intro to two seconds and rewrite the hook" — and the AI director makes the edit for you, on a real timeline you can still see and touch.

The surface is the Trackpad Editor: a timeline with tracks and layers, plus an AI copilot riding shotgun in a chat dock. When you describe a change, the copilot translates it into concrete edit actions and applies them to the live composition. You watch it happen, then iterate. No blank canvas, no manual dragging — just a conversation that produces frames.

How every shot is sourcedThe three paintbrushes

Whatever you ask for, the director builds each shot with one of three tools. Understanding them is the single most useful mental model in VidFarm, because it maps directly to cost — and the director always reaches for the cheapest brush that does the job.

• 1 · Raw clips — cut and remix existing footage (yours, or clipped from a source video). The cheapest brush and the workhorse of short-form. Most viral edits are 90% this.
• 2 · HTML/JS hyperframes — video rendered from HTML, with CSS / GSAP / anime.js motion. Great for kinetic captions, title cards, data callouts, and motion-graphic beats. Cheap, and re-themeable to your brand.
• 3 · Pure AI generation — generated images, video, voice, and music. The most powerful brush and the most expensive — AI video especially — so it's used last, for the shots the other two can't cover.

Manual walkthroughEdit by chatting (the web UI)

Here's the loop in the browser. It's the same whether you're swapping one caption or restructuring a few scenes — you just keep describing changes and the copilot keeps applying them.

1. 1

   Open a forked composition in the editor

   From a template you've forked, you land in the Trackpad Editor →. The timeline shows every layer; the chat dock is on the right (jump straight to it with /editor#chat).

   Open the editor →
2. 2

   Describe the edit in plain English

   Type what you want: "Replace scene 2 with my product demo and shorten the intro to 2 seconds." Be concrete about which layer or scene and what the outcome is — the copilot targets the exact layer and edits it in place.

3. 3

   Watch it apply, then iterate

   The copilot applies each change live on the timeline. If it's not quite right, say so — "make the hook bigger", "cut the last beat" — and keep going until it looks right. Small, specific asks land best.

4. 4

   Know when to go bigger

   For a full multi-scene re-theme, the copilot will proactively recommend handing off to a local coding agent — it can even save a Markdown handoff brief to your My Files →. That's your cue to move the heavy job local (next section).

Web copilot vs local coding agentTwo surfaces, same three paintbrushes

Agentic editing runs on two surfaces. They paint with the same three brushes, but they have very different powers — and knowing which is which saves you a lot of frustration.

The web copilot lives in the editor chat dock → (also Chat → and the Library assistant). It edits the live timeline directly, but it has no shell and no filesystem — and the web editor strips <script> on save, so browser-round-tripped compositions stay declarative (CSS, Ken Burns, transitions, animated captions). It's ideal for short-to-medium edits: caption swaps, a scene or two, a single generation, adding captions, approving and scheduling.

The local coding agent — Claude Code or Codex driving the vidfarm devcli — has a real shell and filesystem, the full edit verb set, JS animation adapters, and free local rendering. It's the one for full rebuilds, batch or scripted work, and render-and-critique loops where the agent grades its own output and tries again.

Claude Code, OpenAI Codex, Gemini CLI, OpencodeDesktop agents — the preferred way

Desktop agents are the preferred way to use VidFarm. A desktop agent is an AI coding agent running on your own machine — Claude Code, OpenAI Codex, Google Gemini CLI, Opencode — driving the vidfarm devcli with a real shell, a real filesystem, and free local rendering.

The best part is the economics: these agents run on the flat-rate subscription you probably already pay — the ~$20/month plan from Anthropic or OpenAI. All the thinking, planning, and editing loops the agent does are covered by that subscription, so you save money and never spend more than your $20/month on the agent side. Your VidFarm wallet is only touched when a job genuinely needs it (an AI generation, a cloud render) — the conversation itself costs you nothing extra.

• Claude Code (Anthropic) — install & docs →. Runs on your Claude subscription.
• OpenAI Codex — install & docs →. Runs on your ChatGPT subscription.
• Google Gemini CLI — install & docs →. Google's open-source terminal agent.
• Opencode Agent — install & docs →. Open-source, works with any model provider.

Where to go next

Now you know the model: chat to edit, three brushes cheapest-first, two surfaces for small vs big. From here, set up your account and try the gentlest first edit: