lc-editor

A local video editor for a computer use agent, made for grokbot. You drive it from an MCP client; there is no timeline UI to click. You send tool calls; ffmpeg renders the file.

Built for short 9:16 reels (1080x1920, 30fps). Captions are stroke-and-shadow text, never a box. Sound is the owner's call, not the editor's: ask before assuming music or natural audio. This version composites multiple layers, applies a small effect pack, expands templates into ordinary timeline items, and can mix owner-imported music with beat sync. Engine rules live in specs/craft.md. Karachi episode structure is an optional preset, not the default.

Needs

• Python 3.11 or newer
• ffmpeg and ffprobe on your PATH

Install

From a clone of this repo:

pipx install .

pip install . works too. pipx keeps the app in its own env and puts lc-editor on your PATH.

From GitHub:

pipx install git+https://github.com/uzairayed/lc-editor.git

Run

lc-editor serve --project ./my-reel

--web starts a page on 127.0.0.1:8765 that can show stills. It only reads.

Cursor

Add this to your MCP config. Use an absolute project path.

{
  "mcpServers": {
    "lc-editor": {
      "command": "lc-editor",
      "args": ["serve", "--project", "/absolute/path/to/project"]
    }
  }
}

On Windows the path looks like C:/Users/you/my-reel.

Tools

Every edit returns { ok, timeline_summary, warnings }. Illegal requests fail out loud. Same op_id twice does not duplicate. review_report must pass before export. Preview stills and the contact sheet are JPEG files on disk, not base64.

The primary track is still the gapless clips list (clip_add, trim, split, reorder, motion, transitions). Two or more sources can share one slot (layout_add). Overlays, look, and music sit beside it:

• Layers: layer_add / layer_update / layer_remove / layer_reorder / layer_transform / layer_keyframe for timed video, image, or text. text_style sets fade, pop, slide, or type-on. effect_add / effect_update / effect_remove attach blur, sharpen, glow, grain, vignette, lut, or color. Raw ffmpeg filter strings are rejected.
• Templates: template_list, template_apply("editorial"|"karachi", bindings=...), template_save. Apply writes ordinary layers and look; the project stays editable.
• Music: default is off. project_set(allow_music=true) is the owner opt-in. Then import_file a local .mp3/.wav/.m4a and music_add. beat_analyze, beat_edit, dry-run beat_sync_preview, then beat_sync_apply. There is no stock catalog; licensing stays with the owner.
• Layouts: layout_add("stack_v"|"stack_h"|"stack_v3"|"grid_2x2", panes=[...]) composites two to four sources into one clip. layout_pane refocuses a cell. layout_clear flattens back to pane 0. A vista or reveal stays full-frame.
• Captions: caption_add still works and syncs to a bound text layer. No box, banner, or scrim.

Optional: project_create(..., preset="karachi") loads series branding. Other reels do not need it.

Capability matrix

Area	In this version
Canvas	1080x1920, 30fps, MCP only
Primary track	Gapless clips, trim/split/reorder, kenburns/punch
Layouts	stack_v, stack_h, stack_v3, grid_2x2. One timeline slot
Layers	Timed video, image, and text overlays with z-order, transform, keyframes
Effects	Registry: blur, sharpen, glow, grain, vignette, lut, color
Text	Stroke-and-shadow only. Motion: fade, pop, slide, type-on
Templates	editorial, karachi; apply expands to ordinary layers
Sound	Natural audio, beds, SFX. Music is opt-in via project_set(allow_music=true) then music_add
Beat sync	beat_analyze, beat_edit, dry-run beat_sync_preview, then beat_sync_apply
Look	One adjustment layer (LUT, grain, vignette) after the cut

Giving this to grokbot

Once the MCP server is in grokbot's config, hand it a prompt like this:

You have an MCP server called lc-editor. It is a video editor you control
entirely through tool calls; ffmpeg does the rendering.

My raw clips are in <folder>. Cut a 9:16 reel of <subject>, 15 to 28 seconds.
Follow the craft rules in specs/craft.md: stroke-and-shadow captions with no
background box, hard cuts on motion.

Before you touch the timeline, ask me what I want for sound: music or natural
audio only. Do not assume either way. If I ask for music, call
project_set(allow_music=true), import my track, music_add it, beat_analyze,
show me beat_sync_preview, then beat_sync_apply only after I confirm.

Work in passes: analyze the footage first, pick the strongest shots, then
match what you see in the keyframes to context/INDEX.md and read only those
scene cards (process in context/HOW.md) before you set durations or roles.
Build the timeline, add a layout when two shots belong on one frame
(two signs, two named places), add layers or a template if needed, then
run review_report and fix every warning before you export. Show me
preview stills at each pass.

Swap the folder, subject, and any series preset (for example preset="karachi") into the prompt as needed.

Scene holds (door as a reveal, tile as a detail, highway as punctuation) live in context/. The agent looks up subjects in context/INDEX.md and reads only the matching cards. That library is editorial judgment. Engine floors in specs/craft.md still win when they conflict.

Not in this version

Multicam, speech-to-captions, a stock music catalog, Drive import, or a browser you edit in.