OpenCutAgent

Edit video on the live Adobe Premiere Pro timeline with Claude.

OpenCutAgent is a Premiere Pro panel + local Node server that cuts your footage the way an editor does: transcribe what is said, find the dead air, the filler, the false starts and the duplicate takes, and remove them from your real sequence, undoably. The AI judgment is Claude itself on your Claude subscription. There is no external LLM API and no per-token AI bill.

Two ways to drive it:

Panel only (default): open the panel in Premiere and click. "Analyze w/ Claude" and "Suggest threshold" run the claude CLI headlessly in the background; you never type a prompt.
Chat (Sync mode): talk to Claude Code ("remove the silences", "cut the duplicate takes") and watch the marks land live in the panel via MCP tools.

Two tabs, two price tags:

Remove Silences is completely free: it measures loudness locally with ffmpeg. No API key, no network, no cost.
Retakes needs a paid transcription: the timeline is transcribed once (cached forever) and Claude then judges the takes. We recommend ElevenLabs Scribe v2 (~$0.22/hour of audio, and only the source ranges actually on your timeline are billed).

Features

Remove Silences tab (free): loudness-based dead-air removal (ffmpeg only, no transcription, no API key). Waveform with draggable threshold, live-recomputing silence zones, pacing presets, margins, and an AI threshold suggestion.
Retakes tab (paid transcription): transcript-based cleanup. Loads the timeline as indexed spoken segments; Claude marks duplicate takes, false starts and filler as Cut; you review, override, protect, then apply.
Soft Apply: instead of deleting, lays colored markers on the timeline (one hue per retake group, green over the suggested keeper) so you can pick final takes by hand.
Fast apply: big cut lists skip in-place razoring and rebuild the tightened sequence via Premiere's own XML round-trip, so a 2-hour timeline with 2000+ cuts applies in seconds, with effects and transforms preserved.
Export transcript: saves the kept speech as a YouTube-ready .srt, with caption times matching the tightened video.
Live sync: the panel follows Premiere's playhead, highlights the segment under it, and clicking a segment seeks the timeline.

How it works

Claude Code ──stdio──► MCP server (Node) ──ws 127.0.0.1──► CEP panel (in Premiere) ──evalScript──► Premiere
             ppro_*     runs ffmpeg + Scribe                auto-reconnects                premiere.jsx
             tools      + headless `claude -p`

The MCP server (server/) hosts a localhost WebSocket bridge for the panel, exposes the timeline as MCP tools, runs ffmpeg + ElevenLabs Scribe, computes every cut list, and spawns the headless claude calls for the panel's AI buttons.
The CEP panel (cep-panel/) runs inside Premiere and executes timeline operations via cep-panel/host/premiere.jsx (the only file that touches Premiere's API).
The skills (.claude/skills/) teach Claude the cut and silence-removal workflows; the same text drives the headless calls.

Premiere's own transcript isn't readable via any API, so transcription runs on the source media (transcribe once, cache, map onto the timeline; a re-edit or reload never re-bills).

For the full picture (reconcile, fast-apply ladder, chunked AI analysis), see docs/ARCHITECTURE.md.

Prerequisites

Adobe Premiere Pro 2024–2026 (uses CEP, which still loads on current 26.x). macOS or Windows.
Node.js ≥ 18 (node -v).
ffmpeg on your PATH (ffmpeg -version):
- macOS: brew install ffmpeg
- Windows: winget install Gyan.FFmpeg (or choco install ffmpeg)
Claude Code installed and signed in (claude on your PATH). Needed for the AI features (headless buttons and Sync mode); scanning, manual thresholds and manual keep/cut work without it.
ElevenLabs API key: optional, only for the transcription features (the Retakes tab, identify_segments, transcript-based remove_silences). Transcription is the one paid piece; we recommend Scribe v2 (the default model, selectable in the panel settings). The key must have the speech_to_text scope. The Remove Silences tab needs no key and costs nothing. Get a key at https://elevenlabs.io.

Too much? Let Claude install everything

If you already have Claude Code, you can skip the entire Install section below. Open a terminal in the folder where you want OpenCutAgent to live, run claude, and paste this prompt. Claude will clone the repo, install the prerequisites it can, wire up the panel and the MCP config, and verify the result:

Install OpenCutAgent on this machine so it is fully ready to use, doing every step yourself and verifying as you go. Steps:

Clone https://github.com/leonardogrig/opencutagent.git into the current directory (skip if already cloned here) and run npm install inside its server/ folder.

Check prerequisites: Node.js >= 18 and ffmpeg on PATH. If ffmpeg is missing, install it (macOS: brew install ffmpeg; Windows: winget install Gyan.FFmpeg). If Node is missing or too old, tell me how to install it and stop.

Enable Adobe CEP developer mode for CSXS.11 and CSXS.12 (macOS: defaults write com.adobe.CSXS.<n> PlayerDebugMode 1; Windows: the HKCU Software\Adobe\CSXS.<n> PlayerDebugMode=1 registry values).

Install the panel: symlink (macOS) or junction/copy (Windows) the repo's cep-panel folder to the user CEP extensions folder as com.opencutagent.panel (macOS: ~/Library/Application Support/Adobe/CEP/extensions/; Windows: %APPDATA%\Adobe\CEP\extensions\). Replace any stale link at that name.

Create .env from .env.example if it doesn't exist. Don't put any API key in it; I can add my ElevenLabs key later from the panel's gear menu.

Create .mcp.json from .mcp.json.example, setting the server path to the ABSOLUTE path of this clone's server/index.js (forward slashes on Windows; no ${VAR} expansion).

Verify: run npm run check and npm test in server/, and confirm the extensions link resolves to the repo's cep-panel.

Finish with a short report of what was installed, anything you could not do, and my two next steps: restart Premiere and open Window > Extensions > OpenCutAgent (the panel auto-starts the server), and optionally add an ElevenLabs key in the panel settings to enable the Retakes tab.

Ask before anything destructive; everything else, just do.

When it finishes, restart Premiere and open Window ▸ Extensions ▸ OpenCutAgent. Done — the manual steps below are the same thing, spelled out.

Install

Replace /path/to/opencutagent (or C:\path\to\opencutagent) with wherever you cloned the repo.

1. Clone + install server dependencies

git clone https://github.com/leonardogrig/opencutagent.git
cd opencutagent/server
npm install

2. API key (optional)

cd ..                  # back to the opencutagent root
cp .env.example .env   # Windows: copy .env.example .env

Edit .env and set ELEVENLABS_API_KEY=... only if you want the transcription features. .env also holds optional knobs (port, cache dir, AI model/effort defaults, fast-apply thresholds), all documented inline.

3. Enable CEP developer mode + install the panel

Premiere only loads unsigned dev panels with developer mode on (CEP 11 = Premiere 2024, CEP 12 = 2025/2026).

macOS

defaults write com.adobe.CSXS.11 PlayerDebugMode 1
defaults write com.adobe.CSXS.12 PlayerDebugMode 1

# link the panel into Premiere's extensions folder
mkdir -p ~/Library/Application\ Support/Adobe/CEP/extensions
ln -s /path/to/opencutagent/cep-panel \
      ~/Library/Application\ Support/Adobe/CEP/extensions/com.opencutagent.panel

Windows (PowerShell)

reg add HKCU\Software\Adobe\CSXS.11 /v PlayerDebugMode /t REG_SZ /d 1 /f
reg add HKCU\Software\Adobe\CSXS.12 /v PlayerDebugMode /t REG_SZ /d 1 /f

# junction the panel folder (run PowerShell as Administrator)
New-Item -ItemType Junction `
  -Path "$env:APPDATA\Adobe\CEP\extensions\com.opencutagent.panel" `
  -Target "C:\path\to\opencutagent\cep-panel"

No admin rights? Just copy cep-panel to %APPDATA%\Adobe\CEP\extensions\com.opencutagent.panel instead of the junction (but re-copy after any panel update).

Restart Premiere, then open the panel: Window ▸ Extensions ▸ OpenCutAgent. That's enough for panel-only use: the panel auto-starts the server when you open it.

4. Point Claude Code at the server (for Sync / chat use)

Copy the example MCP config and set the path to your clone:

cp .mcp.json.example .mcp.json

{
  "mcpServers": {
    "premiere": {
      "type": "stdio",
      "command": "node",
      "args": ["/absolute/path/to/opencutagent/server/index.js"],
      "timeout": 600000
    }
  }
}

Use an absolute path; Claude Code does not expand ${VAR}s in .mcp.json. On Windows use a forward-slashed absolute path, e.g. C:/path/to/opencutagent/server/index.js.

Then launch Claude Code from the project folder so it loads .mcp.json and the bundled skills, and approve the server when prompted:

cd /path/to/opencutagent
claude

(Alternative: register globally with claude mcp add premiere --scope user -- node /path/to/opencutagent/server/index.js, and symlink the skills into ~/.claude/skills/ so they load from any folder.)

Who starts the server (one server, three launchers)

The panel is only a WebSocket client; everything runs through one local server (server/index.js, port 3001). It can start three collision-safe ways:

Opening the panel auto-starts it (the panel spawns it only after a failed connect, so it never fights an existing one). This is all you need for panel-only use.
Claude Code spawns it via .mcp.json when a session is open in the project. For Sync mode, start the claude session first so its server owns the port; the panel then connects to it.
Manually: npm start from the repo root (stop with npm stop), or double-click start-opencutagent.command on macOS.

To smoke-test the server by itself (no Premiere needed):

cd /path/to/opencutagent/server
npm run smoke   # boots the server, lists the tools over stdio

After changing server code, restart the Claude session (or /mcp → reconnect premiere); there is no hot-reload.

Use it: Remove Silences (loudness-based, no key)

Open a project + sequence, open the panel, click Scan Audio. The server measures each clip's loudness (ffmpeg, cached per source) and the panel draws the waveform with silences highlighted and a draggable Noise Threshold line.
Tune live: drag the threshold, edit the ms fields (min silence length, keep-talk, margins), or pick a pacing preset. The zones recompute instantly. The meter is normalized to the recording's own peak, so threshold values behave like AutoCut/TimeBolt's.
Suggest threshold (optional): Claude reads the measured loudness stats and picks a threshold that fits the recording (headless by default; via your chat in Sync mode).
Pick the Silence Management mode: Remove (ripple, close gaps), Keep gaps (lift), Mute, or Keep.
Remove Silences. Undo restores the timeline.

Use it: Retakes (transcript-based)

Load segments: transcribes the timeline (ElevenLabs Scribe, cached; only the source ranges actually used on the timeline are billed) and lists indexed spoken segments, each starting as Keep. Clips with no speech are auto-marked Cut.
Analyze w/ Claude: Claude reads the transcript and marks duplicate takes, false starts and filler as Cut (keeping the most complete pass of each beat). Long timelines are analyzed in overlapping chunks for reliability.
Review: click a segment to expand; flip Keep ⇄ Cut, or Protect it so nothing ever cuts it. Overrides show a Manual badge. The list follows the playhead; clicking a segment's time seeks Premiere.
Optional extras before applying:
- Soft Apply lays colored markers instead of deleting (green = suggested keeper); Clear markers removes only OpenCutAgent's markers.
- Remove excess (keep speech only) also trims the dead air inside kept segments.
- Export transcript saves the kept speech as an .srt.
Apply All. "Remove gaps when applying" picks ripple vs lift. Large ripple applies (100+ cuts) build a new sequence named <sequence> - tightened via XML round-trip (fast, effects preserved); the original sequence is untouched, so "undo" there is just deleting the new one. Smaller applies edit in place and support Undo last apply / Cmd+Z.

Or just ask in chat (Sync mode): "analyze the retakes on my timeline and cut the duplicates", "remove the silences", "what's on the timeline?", "trim V1.2 so it ends at 00:00:18:00". The premiere-edit skill runs a read → propose → confirm → execute → verify loop and previews destructive cuts before applying.

The AI: Claude, two transports

The "Sync with Claude Code" toggle (in the panel's ✦ settings popover) picks who runs the analysis:

OFF (default, headless): the AI buttons spawn claude -p in the background on your Claude subscription (keychain login; a stray ANTHROPIC_API_KEY is scrubbed so nothing silently bills). Model + effort come from the settings popover.
ON (Sync): the buttons defer to your open Claude Code chat, where Claude drives the ppro_* MCP tools and pushes results into the panel live.

Both modes share the same server state and render identically. The headless call is a pure judgment oracle: it gets segment data and returns JSON decisions; it cannot touch Premiere, your files, or the network.

MCP tools

Tool	What it does
`ppro_get_timeline_state`	Read the sequence: clips (ids like `V1.2`), source paths, timecodes, gaps, `revision`.
`ppro_identify_segments`	Transcribe a clip's source (Scribe, cached) → phrases / silences / fillers at timeline timecodes.
`ppro_trim_clip`	Set absolute new edges for one clip (idempotent).
`ppro_remove_gaps`	Ripple-close empty gaps.
`ppro_remove_silences`	Transcribe → cut list (silences + fillers) → ripple-delete. Use `dry_run: true` first.
`ppro_analyze_audio_levels`	Measure timeline loudness (ffmpeg, no transcription) → noise-floor/speech stats + suggested threshold.
`ppro_remove_silences_by_level`	Loudness-based silence removal from chat: threshold → cut list → ripple/lift/mute. `dry_run` first.
`ppro_get_retake_segments`	Transcribe into indexed segments for Claude to analyze for retakes/duplicates.
`ppro_mark_retakes`	Record keep/cut decisions; pushes them live to the panel.
`ppro_apply_retakes`	Apply the current marks (ripple/lift-delete the Cut segments).
`ppro_run_script`	Escape hatch: run arbitrary ExtendScript.

Verify the install

cd /path/to/opencutagent/server
npm run check     # syntax-check every module
npm test          # unit + feature + smoke tests (no Premiere or API key needed)
npm run smoke     # start the server + list tools over stdio

Then open the panel (it should show Connected) and click Scan Audio, or ask Claude Code "what's on the timeline?".

Troubleshooting

Panel says "Waiting for server…". Nothing is listening on the port. Reopen the panel (it auto-starts the server), run node server/index.js yourself, or open a claude session in the project. If auto-start says "Node not found", install Node or start the server manually once.
Panel not in the Extensions menu. macOS: re-check the defaults write … PlayerDebugMode 1 commands and that the symlink points at cep-panel/. Windows: re-check the CSXS.11/CSXS.12 registry keys and the junction/copy at %APPDATA%\Adobe\CEP\extensions\com.opencutagent.panel. Restart Premiere.
"Unknown RPC method …" after updating the code. A server running the old code still owns the port (commonly from a second Claude Code window in the project). Keep one window, then /mcp → reconnect premiere (or restart the session). If a stray server lingers: find it with lsof -nP -iTCP:3001 -sTCP:LISTEN (Windows: netstat -ano | findstr :3001) and kill that PID.
"EvalScript error." premiere.jsx didn't load; close and reopen the OpenCutAgent panel.
Port 3001 busy. Set PREMIERE_BRIDGE_PORT in .env; the panel auto-reads the negotiated port from ~/.editagent/bridge-port.
Transcription fails / Retakes tab errors. Confirm ELEVENLABS_API_KEY is set, has the speech_to_text scope, and ffmpeg -version works. (The Remove Silences tab needs neither a key nor network.)
A "Translation Report" alert during a big apply. Benign: Premiere logs source-interpretation entries it re-derives on import. The report file lands in .cache/rebuild/ if you want to read it.

Good to know

OpenCutAgent is built for the common case: talking-head footage on a normal A/V timeline. Two automatic behaviors worth knowing about:

Big applies build a new sequence. Large ripple applies rebuild the tightened cut via Premiere's own XML round-trip (seconds instead of minutes, with your effects and transforms preserved); the original sequence is left untouched as a backup.
The fast path steps aside when it must. Timelines with titles/graphics or speed-changed clips are applied by the (slower) in-place razor path instead. Everything still applies either way.

The transcription engine is pluggable (Scribe v2 today); the engine interface in server/transcription/transcribe.js accepts a Deepgram/Whisper drop-in.

Docs

docs/ARCHITECTURE.md: how the pieces talk, the fast-apply ladder, the design rules.
CONTRIBUTING.md: dev setup, tests, reload matrix, platform gotchas.
cep-panel/DESIGN.md: the panel's design system.

License

MIT with Commons Clause. In short: free to use, modify, and share for yourself or your team, but you may not sell OpenCutAgent or offer it (or a product/service substantially built on it) commercially. That right is reserved by the author.

Name		Name	Last commit message	Last commit date
Latest commit History 4 Commits
.claude/skills		.claude/skills
.github/workflows		.github/workflows
cep-panel		cep-panel
docs		docs
server		server
.env.example		.env.example
.gitignore		.gitignore
.mcp.json.example		.mcp.json.example
CLAUDE.md		CLAUDE.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.md		README.md
package.json		package.json
start-opencutagent.command		start-opencutagent.command

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

OpenCutAgent

Features

How it works

Prerequisites

Too much? Let Claude install everything

Install

1. Clone + install server dependencies

2. API key (optional)

3. Enable CEP developer mode + install the panel

4. Point Claude Code at the server (for Sync / chat use)

Who starts the server (one server, three launchers)

Use it: Remove Silences (loudness-based, no key)

Use it: Retakes (transcript-based)

The AI: Claude, two transports

MCP tools

Verify the install

Troubleshooting

Good to know

Docs

License

About

Releases

Packages

Contributors

Languages

Folders and files

Latest commit

History

Repository files navigation

OpenCutAgent

Features

How it works

Prerequisites

Too much? Let Claude install everything

Install

1. Clone + install server dependencies

2. API key (optional)

3. Enable CEP developer mode + install the panel

4. Point Claude Code at the server (for Sync / chat use)

Who starts the server (one server, three launchers)

Use it: Remove Silences (loudness-based, no key)

Use it: Retakes (transcript-based)

The AI: Claude, two transports

MCP tools

Verify the install

Troubleshooting

Good to know

Docs

License

About

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages