<div align="center"> <img src="docs/assets/logo.png" width="104" height="104" alt="Lumen logo" />

Lumen — the open-source AI video editor

Edit video with an AI that can see, hear and cut.

_{Vibecoded with <b>Claude Opus 5.5</b> by <a href="https://github.com/Mas-inx"><b>Mas-Inx</b></a> at <a href="https://async.tech/"><b>AsyncTech</b></a>}

Lumen is a desktop video editor where the AI isn't a chatbot bolted onto a timeline. It watches your edit frame by frame, listens to the mix, reads what's said — and makes real, undoable edits with the same tools you use. Bring your own brain: Claude Code, Codex, or a model with your OpenAI, Anthropic, Gemini, OpenRouter or local Ollama / LM Studio key.

 

<img src="docs/screenshots/editor.jpg" alt="The Lumen AI video editor: a chrome 3D title over a Blender-rendered background in the preview, the media library of generated clips, a timeline with captions, titles, overlays, three backgrounds, narration and music, and the Copilot panel showing its steps" /> </div>

Why Lumen

• An AI that edits, not just chats. Every edit in Lumen is a typed command, and the AI gets all of them — plus tools to look and listen. More than 100 tools in all, each change a normal undo step. See them all →
• It sees what you'll export. When the AI checks a frame, it's rendered exactly as the export will be: every track, title, effect and transition. It can watch the whole edit as a contact sheet, look inside footage, and hear loudness, silences and speech.
• Bring your own brain. Use the Claude Code or Codex you already pay for, an API key from any major provider, or a local model. Lumen is also an MCP server, so any agent can drive it.
• Local-first. Media is edited in place and never uploaded. Speech-to-text runs on your GPU with Whisper. No account, no watermark, no telemetry.
• A real editor underneath. A magnetic timeline, keyframes, 3D titles and transitions, colour looks, frame-exact export with hardware encoding — so you can finish by hand whatever the AI starts.

It's an open-source alternative to editors like CapCut, Filmora and Descript — built so an AI can do anything you can.

<div align="center"> <img src="docs/screenshots/teaser.webp" width="720" alt="An animated teaser made entirely in Lumen: a chrome 3D title spinning in over Blender-rendered glossy shapes, a caption, and a 3D cube transition into the next shot" /> <br /> _{Made in Lumen, frame for frame: Blender backgrounds and a chrome 3D title, HyperFrames motion graphics, Whisper captions, and a 3D cube transition the Copilot added and checked.} </div>

Contents

• Features — editing · Copilot · smart edits · generation · export · integrations
• Download · Build from source · Documentation · Roadmap · Contributing · Credits

Features

A real editor

<img src="docs/screenshots/timeline.jpg" alt="Lumen's multi-track timeline: captions, a closing 3D title, overlay renders (3D title, lower third, kinetic type), three Blender backgrounds on the magnetic main track, narration and a ducked music bed, with markers on the ruler" />

• Magnetic main track that keeps your story edge to edge, with free tracks above and below for B-roll, overlays, titles, captions, voice, music and sound effects.
• Split, trim, ripple, slide into gaps, snapping, markers, speed and reverse, and keyframes on every property:
  • 28 easing curves per keyframe, or your own cubic-bezier;
  • one-click motion presets baked into keyframes: Ken Burns, push-ins, pans, pops, slides, and a punch that lands on the beat;
  • keyframes you drag right on the timeline.
• Render previews like Premiere and Filmora:
  • a bar under the ruler shows what may stutter and what's rendered;
  • Enter renders in to out, or Lumen renders heavy stretches while you're idle;
  • render at Full, 1/2, 1/4 or 1/8 resolution, sharpened back up on playback.
• Smooth on modest computers. Nothing redraws while idle, filmstrips and waveforms are painted once and reused, and the preview drops resolution when it has to.
• Pro editing tools: roll, slip and slide; ripple trim and insert or overwrite on any track; in and out points to play, lift, extract or export a stretch; copy, cut, paste and paste attributes; crop with rounded corners; freeze frames and speed ramps; detached sound for J- and L-cuts; groups; and solo and reorder for tracks.
• 2D and 3D transitions (cube, flip, door, page curl, shatter…) placed before, on or after a cut, effects, colour looks and grading, and extruded 3D titles rendered live.
• Real colour grading on the GPU: tone curves, lift / gamma / gain wheels, HSL by colour range, imported .cube LUTs, and scopes (waveform, RGB parade, vectorscope, histogram) beside the program monitor.
• Compositing: chroma and luma keys with an eyedropper and spill suppression; feathered rectangle and ellipse masks you drag on the canvas (on adjustment layers too, to grade just part of the frame); outlined titles in any font — imported, or picked from those installed on your computer, and embedded in the project.
• Text behind subject: on-device AI cuts the person (or, with a GPU, any subject) out of a shot. Slip a title between them and the background, remove the background, or remove the person. One click, one undo step, frame-exact on export.
• Sound that drives the picture: a Web Audio mix with per-clip volume, fades and keyframes, Studio voice clean-up, on-device noise reduction and 15 built-in sound effects.
• A real mixer: a channel per track and a master bus, each with a fader, pan, a draggable EQ curve, a compressor and a limiter, with live meters. Transitions crossfade the sound too, and exports can be loudness-normalized to −14 LUFS for YouTube and Spotify, −16 for Apple, or broadcast's −23 and −24.
• Import MP4, MOV, WebM, MKV, MP3, WAV, FLAC, PNG, JPEG, WebP, GIF and more — by dialog, folder or drag and drop. Files stay where they are; missing ones can be relinked. Select many at once (Ctrl/Shift-click or drag a box) to add, transcribe or remove them together.
• .lumen project files that survive moving folders between drives, crash recovery, and undo for everything. Ctrl+K reaches every command.
• Built for long, heavy projects: proxies for 4K footage (made automatically in the background; exports always use the originals), hour-long recordings that stream instead of filling memory, version history (every save, plus a snapshot every ten minutes while you edit, restorable any time) and Collect project to gather a project and all its media into one folder.
• Multiple timelines and nesting: as many timelines per project as you like — another cut, or a vertical copy reframed in one click — and nest any clips into a timeline of their own to grade, speed up or cut them as one piece (double-click to step inside; picture and sound both play through).
• Multicam: pick the camera recordings and Lumen lines them up by their sound (to about a millisecond), then press 1–9 while it plays to cut between angles, with every camera live in a strip under the preview.
• Reads the footage: scene detection splits or marks a clip at every shot change; stabilization smooths away camera shake with just enough zoom; motion tracking pins a title, sticker or mask to something moving — including a blur that follows a face.

The Copilot sees, hears and edits

<table> <tr> <td width="46%"><img src="docs/screenshots/copilot.jpg" alt="The Lumen Copilot reading the project, watching a contact sheet, applying a batch of edits and checking a frame it rendered" /></td> <td>

Ask in plain words — or talk to it. The Copilot plans, uses Lumen's tools, and checks its own work:

| You ask | It… | | --- | --- | | “Remove the long pauses and add captions” | transcribes on your GPU, cuts every silence across the whole timeline (captions and B-roll stay in sync), lays timed captions | | “Make the intro punchier” | watches the edit, adds a 3D transition and a glow in one undo step, renders the frame to check it | | “Is the music too loud under my voice?” | measures loudness per track in LUFS and ducks the music under speech | | “Make a vertical version for Reels” | reframes the canvas and every layer to 9:16 | | “Find where I talk about the mountains” | searches inside transcripts and jumps there | | “Put the title behind me” | cuts you out of the shot on your computer and slips the title between you and the background | | “Cut the B-roll to the music” | finds the beat and lays the clips out on it |

You see every step, including the frames it looked at, and Undo all reverts a whole answer.

</td> </tr> </table>

What the AI sees when it watches your edit — one call to get_contact_sheet:

<img src="docs/screenshots/contact-sheet.jpg" alt="A contact sheet the Lumen AI generated of the whole edit: labelled frames of 3D backgrounds, the chrome title, the lower third and captions" />

Brains: your own Claude Code or Codex (Lumen runs them headless, with only Lumen's tools), or a model with your key: Anthropic, OpenAI, Google Gemini, OpenRouter, OpenCode Zen, OpenCode Go, Ollama, LM Studio or any OpenAI-compatible endpoint. Pick the model and the effort — how hard it thinks, from Low to Max (Ultra on Codex) — for every one of them, Claude Code's Fable, Opus, Sonnet and Haiku and Codex's own model list included. Keys are encrypted by Windows (DPAPI) and never leave the app's main process.

• Chats per project. Start new chats, switch between them, and pick up where you left off after a restart.
• Nothing happens out of sight. The chat shows the model's thinking, each step while it is written, run and finished, and a line with a clock for what it is doing right now. Open any step to see what was sent and what came back.
• **Queue a