Control Adobe Premiere Pro from Claude, Cursor, or any MCP-compatible AI client — real timeline editing, not just project metadata.
Developer: CaYaDev · cayadev.com
Important
To install: go to Releases, download PPMCP-Setup-x.x.x.zip, extract it, then double-click Setup.bat. That single file runs the whole installer.
| Language | File |
|---|---|
| English | README.md (this file) |
| Turkish | README.tr.md |
| Spanish | README.es.md |
| German | README.de.md |
| French | README.fr.md |
| Japanese | README.ja.md |
| Chinese (Simplified) | README.zh-CN.md |
Full step-by-step guide: INSTALL.md
(paths, UXP Developer Tool download, Claude / Claude Code / Cursor, Premiere panel)
- Open GitHub Releases.
- Download
PPMCP-Setup-x.x.x.zip. - Extract the folder → double-click
Setup.bat. - The wizard shows install folder, version, and optional CEP Text Bridge.
- If PPMCP is already installed, Setup offers:
- Update / reinstall to this version
- Uninstall completely
- When finished, open the personalized guide (real full paths on your PC):
%APPDATA%\PPMCP\HOW-TO-USE.txt%APPDATA%\PPMCP\HOW-TO-CONNECT.txt%APPDATA%\PPMCP\mcp-config-snippet.json
Setup is pure PowerShell (built into Windows) — no third-party installer, commercial-safe.
The ZIP bundles portable Node.js — you do not need to install Node yourself.
You do need Adobe UXP Developer Tool once (free) to load the Premiere panel:
- https://developer.adobe.com/premiere-pro/uxp/plugins/distribution/install/
- https://developer.adobe.com/photoshop/uxp/2021/devtool/installation/
- Search: Adobe UXP Developer Tool download
npm run release:winNeeds only Node.js + npm. Output: dist-release/PPMCP-Setup-*.zip
git clone https://github.com/CaYatur/PremiereProMCP.git
cd PremiereProMCP
npm install
npm run build
npm run dev:bridgeOr (Node already installed): installer\Setup.bat or installer\install.bat
Load plugin/manifest.json in UXP Developer Tool. Point MCP at server/dist/index.js.
Details: INSTALL.md.
PPMCP is a local stdio MCP server — a Node process on your own PC that your AI client launches and talks to over stdin/stdout. There is no URL and no port to enter.
Warning
Claude's Settings → Connectors → Add custom connector asks for a Remote MCP server URL. That dialog is for hosted servers and will not work for PPMCP. Use the client's local config file (or claude mcp add) instead.
Every client needs the same two values, and Setup already wrote them — with your real paths — into %APPDATA%\PPMCP\HOW-TO-CONNECT.txt and mcp-config-snippet.json:
| Value | Default |
|---|---|
command |
%LOCALAPPDATA%\PPMCP\node\node.exe |
args[0] |
%LOCALAPPDATA%\PPMCP\server\dist\index.js |
Escaping: inside a
.jsonfile every backslash must be doubled (C:\\Users\\You\\...); on a command line it must not be. This is the most common connection failure.
Claude Desktop — Setup writes this for you. By hand: edit %APPDATA%\Claude\claude_desktop_config.json (create it if missing), merge into the existing "mcpServers" object, then fully quit and reopen the app — closing the window is not enough.
{
"mcpServers": {
"premiere-pro": {
"command": "C:\\Users\\You\\AppData\\Local\\PPMCP\\node\\node.exe",
"args": ["C:\\Users\\You\\AppData\\Local\\PPMCP\\server\\dist\\index.js"]
}
}
}Claude Code (CLI) — --scope user makes it available in every project; check with claude mcp list.
claude mcp add premiere-pro --scope user -- "C:\Users\You\AppData\Local\PPMCP\node\node.exe" "C:\Users\You\AppData\Local\PPMCP\server\dist\index.js"Cursor — same JSON as Claude Desktop, in %USERPROFILE%\.cursor\mcp.json (global) or .cursor\mcp.json (this project only).
VS Code / GitHub Copilot agent mode — the one client with a different shape: the key is servers, not mcpServers, and it needs an explicit "type": "stdio". Put it in .vscode/mcp.json, or run the MCP: Open User Configuration command.
{
"servers": {
"premiere-pro": {
"type": "stdio",
"command": "C:\\Users\\You\\AppData\\Local\\PPMCP\\node\\node.exe",
"args": ["C:\\Users\\You\\AppData\\Local\\PPMCP\\server\\dist\\index.js"]
}
}
}Windsurf — %USERPROFILE%\.codeium\windsurf\mcp_config.json (or Cascade panel → MCP icon → Configure), mcpServers shape, then reload Windsurf.
Any other MCP client — Cline, Roo Code, Continue, Zed, LM Studio, JetBrains AI, Gemini CLI, Codex CLI, or your own host: give it transport stdio (some clients call it "local", "command", or "process"), the Node path as the command, and the server path as the single argument. Nearly all of them use the mcpServers shape above; VS Code's servers + "type": "stdio" is the only common variant. A client that offers only a remote URL field cannot run PPMCP.
Test without any client — this should print a startup line and then wait (Ctrl+C to stop). If it does, the server is fine and the problem is your client's config:
"C:\Users\You\AppData\Local\PPMCP\node\node.exe" "C:\Users\You\AppData\Local\PPMCP\server\dist\index.js"Building from source instead of the Setup ZIP? Run npm run build, then use node (system Node 18+) as the command and the absolute path to your clone's server/dist/index.js as the argument.
All 277 tools are loaded. A profile only decides how many are registered when the session starts, because putting 277 schemas in front of a model costs thousands of tokens and measurably worsens tool selection.
PPMCP_PROFILE |
Registered at start | Use when |
|---|---|---|
core |
19 | Small/cheap models; edit_bootstrap → edit_auto → edit_verify only |
standard (default) |
109 | Live-verified tools plus the atomics a real cut uses |
full |
277 | Entire catalog resident from the first turn, as in 1.0.x |
The model can widen this itself, mid-session. If it decides it needs more, it calls tool_profile — no restart, no config edit, no asking you:
tool_profile { profile: "full" } // register all 277
tool_profile { category: "color" } // register one area, additive
tool_profile { enable: ["clip_reverse"] }
The server registers them and emits notifications/tools/list_changed, so your client refreshes its own tool list. Alongside that, tool_search finds tools that are not registered and tool_invoke runs any tool whether registered or not — so nothing is ever unreachable, whichever profile you pick.
"env": { "PPMCP_PROFILE": "standard" }Full per-client walkthrough and a troubleshooting table: INSTALL.md.
PPMCP is a Model Context Protocol server that connects an AI agent to a live Adobe Premiere Pro session. The model can:
- Create sequences and assemble media
- Cut, overwrite, and trim clips
- Place titles, shapes, and motion graphics
- Mix audio (SFX, music beds) with safe gain defaults
- Grade, transitions, quality pass, export, screenshots
UXP-first core editing. Optional CEP bridge only for editable MOGRT titles (PNG titles work without it).
MCP Client (Claude / Cursor / ...)
<-> stdio
MCP Server (Node)
<-> WebSocket :8265
Bridge / Relay
<->
UXP Plugin (inside Premiere Pro)
| Requirement | Notes |
|---|---|
| Windows | Primary tested platform |
| Adobe Premiere Pro | UXP-capable (e.g. 25 / 26+) |
| UXP Developer Tool | Required to load the panel (download) |
| MCP client | Claude Desktop, Cursor, Claude Code, etc. |
| Node.js 18+ | Only if you install from source (EXE includes portable Node) |
Optional: ffmpeg / ffprobe for media analysis scripts.
- Bridge running (Desktop PPMCP Bridge if needed)
- UXP Developer Tool → Add Plugin →
...\plugin\manifest.json→ Load - Premiere panel shows Active (footer: CaYaDev · cayadev.com)
- MCP client uses paths from
%APPDATA%\PPMCP\mcp-config-snippet.json - Call
edit_bootstrap
Agent workflow: docs/AGENT_USAGE.md · skill/SKILL.md
- Import media, create sequences, set active sequence by exact name
- Place clips with overwrite; trim in/out
- Markers; sequence screenshots for visual QA
text_write(PNG always; editable MOGRT with optional CEP)- Shapes, colors, Motion keyframes (scale / position / rotation / opacity)
- SFX and music beds; dB API (0 dB = unity)
- Scoped gain (does not mass-overwrite the user’s mix)
quality_pass(batched for large timelines), transitions, export
| Feature | Why |
|---|---|
| Rate limits | Protects Premiere from tool spam (RATE_LIMITED) |
| Checkpoints | Snapshot / restore project before risky edits |
Playbooks / edit_run |
Multi-step edits without thrash |
| Agent docs | Shared rules for different AI models |
Rough cuts · cinematic SFX · music-rhythm cuts · ad-style motion cards · QA before export
277 MCP tools across ~20 categories. Since 1.1.0 the server registers a profile (109 by default) rather than the whole catalog — the rest stay one call away via tool_search → tool_schema → tool_invoke, and PPMCP_PROFILE=full restores the old surface. The catalog spans project, sequence, track, clip, transitions, effects — including 52 one-shot dedicated effect/audio/transition shortcuts — color/Lumetri, audio, text/titles/shapes, markers/metadata, multicam, proxy/media, export, analysis, batch, selection/system, checkpoints, agent-orchestration/edit-pipeline, plus ~22 high-level workflow tools). Most of that surface maps to a real, documented method in Adobe's own @adobe/premierepro UXP API. Real end-to-end sessions have now exercised a much wider slice of it than the ~48s smoke sequence below; the issues surfaced so far are the ones flagged below (see "Still broken" and "Lower-confidence claims") — see docs/FEATURES.md for the full tool-by-tool verification tier (type-verified / live-verified / verified-composed / known-broken).
What holds up well under real, end-to-end testing (a real ~48s multi-track sequence built from scratch, video + 4 audio tracks, transitions, gain, keyframed fades, markers, title, screenshot, save): sequence/project creation, clip_overwrite, trim, roll/slip/slide, split, ripple delete, shape add + position + fill color, text_write's PNG fallback path, listing effects/transitions, gain/dB control, and project save/screenshot. clip_append is now confirmed working in a real session — it appends clips in the correct order (it previously failed with "Script action failed to execute"; the shared-retry fix held up live). sequence_set_in_out is confirmed working too (re-tested 2026-07-11) — it set the in/out points via "via": "sequence.createSetInPointAction + sequence.createSetOutPointAction", confirming the 1.0.1 root-cause fix (the factory is on the Sequence object, not SequenceEditor). Multi-step edits (roll/slip/slide and composite workflow tools) are committed through Premiere's Project.executeTransaction() — several primitive actions run as one atomic unit, so a failure partway through doesn't leave the timeline half-edited. That transaction design has been the most reliable part of the whole plugin.
clip_insert and marker_add are more reliable than earlier testing suggested. Both now retry ~10-15 internal variants (different track-index/limit-shift/marker-type combinations) inside the plugin before giving up — in the latest real session, clip_insert succeeded on every one of ~19 video/audio clips (via one of those retry variants), and marker_add added 7 markers with no reported error. This retry loop is also part of why individual calls can feel slow. We don't yet know for certain whether that marker_add run went through the native path or the virtual-marker fallback (both return the same success message) — treat marker_add as "probably fine, unconfirmed which path," not "known broken."
Fix applied, pending a live re-test (not yet re-confirmed):
| Tool | What was wrong | What changed |
|---|---|---|
app_get_version |
Always returned null — the code read version off the Application class, but per Adobe's ppro_reference version is an instance property (Promise<string>, 25.6+), so it never resolved |
Now reads the version from the UXP host object (require("uxp").host.version), the build-agnostic path every UXP app exposes; returns { version, host, uxpVersion }. Fixed in 1.0.2 (ppro.Application.version kept only as a cheap guard) |
media_get_info / media_analyze_file_info |
Never returned duration, resolution, or frame rate | Now opportunistically probes a few extra fields (getDuration, getFrameSize, width/height, getFrameRate), feature-detected — returns them if this Premiere build happens to expose any, undefined otherwise (no worse than before) |
Still broken — confirmed Adobe platform limitation, not a plugin bug:
| Tool | Issue | Use instead |
|---|---|---|
track_add / track_add_video / track_add_audio |
The Premiere UXP API exposes no method to add an empty track — verified 2026-07-11 against Adobe's official reference: neither the Sequence class nor SequenceEditor has any addTrack/addVideoTrack/addAudioTrack/createAddTrackAction. This is missing from Premiere itself; no plugin-side code can add it |
Choose the track count when you create the sequence (sequence_create, or a sequence preset that already has enough tracks). There is no way to add tracks afterward — and the earlier "drop a clip at a higher track index to auto-create tracks" idea does not work here either: it was tested 2026-07-11 and failed with "[INTERNAL_ERROR] BE: An invalid track index was passed to the sequence" |
sequence_create / sequence_create_from_media track count |
You get the preset's fixed track count (≈3 video + 3–4 audio) and — because of the track_add limitation above — can't add tracks afterward, by any means |
Plan the track count at creation (a preset with enough tracks). Inserting clips beyond the existing track count fails rather than growing the track list |
Lower-confidence claims (from earlier static code analysis, not re-verified in this session): text_set_content (editing existing MOGRT text) and shape_set_size (exact pixel sizing) were previously found broken via code-level probes — see docs/FEATURES.md for the detail — but given clip_insert turned out to be more fixed than that same style of analysis suggested, treat these as "worth re-checking live," not gospel.
Also found and fixed: a fade/keyframe ordering footgun, not a bug. workflow_audio_fade/workflow_fade_clip calculate fade keyframes from the clip's current start/end at the moment you call them — correct in isolation, but if you trim the clip's length after adding fades, the old fade-out keyframe can end up past the new out-point and silently stop applying (fade-in still works, fade-out doesn't). Both tools' descriptions now explicitly warn the model to trim to final length first.
This section gets updated as real Premiere sessions surface more ground truth — static code analysis alone has already been wrong once here (clip_insert).
| Path | Role |
|---|---|
server/ |
MCP tools, rate limit, checkpoints |
bridge/ |
WebSocket relay (:8265) |
plugin/ |
UXP panel in Premiere |
installer/ |
Windows PowerShell Setup (Setup.bat) + release packager |
docs/ |
Architecture, agent usage, roadmap |
skill/ |
Compact AI skill |
scripts/ |
Smoke runs + the CI catalog/profile checks |
See docs/ARCHITECTURE.md, docs/FEATURES.md and docs/ROADMAP.md.
- Soft gap between plugin calls ~100 ms
- Min gap between MCP tools ~220 ms
- High per-minute caps; too fast →
RATE_LIMITED
npm ci # restores workspace links; plain npm install also works
npm run build
npm run typecheck
node scripts/check-catalog.mjs # catalog integrity, no Premiere needed
node scripts/check-profiles.mjs # tool-profile integrity
npm run dev:bridge
npm run release:win # build Setup ZIP (PowerShell wizard + portable Node)CI runs build + typecheck + both checks on Windows and Linux (Node 20 / 22).
Anything needing a live Premiere Pro session stays in scripts/smoke-*.mjs
and is run by hand.
MIT · CaYaDev · cayadev.com
Not affiliated with Adobe. Keep project backups / checkpoints.
