Premiere-MCP-IMT + SKILLS — Premiere Pro for AI agents

io.github.Immersive-Media-TechnologiesPublished by GitHub account Immersive-Media-Technologiesv1.3.1Updated Nov 7, 2026

Premiere Pro for AI agents: every UI language, masks and speed keys by sight, skills, generative AI

VerifiedSTDIODesktop onlyMedia & DesignAI & ML

Overview

AI-generated overview

Lets an AI assistant work inside a running Adobe Premiere Pro: inspect sequences, cut and trim clips, apply effects and keyframes, mask, caption and export.

What you can do
  • Inspect the open project and sequences, listing tracks, clips, times and speed, and export any frame as an image for the model to look at.
  • Cut, trim, move and ripple clips, add markers and captions, reframe, and save the project.
  • Discover and run roughly 440 Premiere operations through a small set of tools, with edits read back from Premiere to confirm them.
  • Draw Object Masks and place speed keyframes by sight on macOS, using real mouse events on Premiere's own UI.
  • Read a bundled skill library covering Premiere's AI tools, recipes and known limitations.
When to use it
Use it when you edit in Premiere Pro and want an assistant to inspect, assemble or clean up sequences in plain language, including on a non-English Premiere UI. It suits tasks like tightening interview pauses, making social aspect-ratio versions, blurring a face, or cutting to a music beat.
Requirements
Premiere Pro 2024–2026 running on macOS or Windows, plus a small CEP connector extension installed inside Premiere (the agent can install it with consent, then Premiere must be restarted). Node.js 20.19+ for the npm route; the Claude Desktop extension needs no Node. FFmpeg for loudness, silence, beat and scene-change analysis. Optional environment variables control read-only mode, raw scripting, the tool surface and the connector mailbox.
Before you install
It edits real Premiere projects and scripted undo covers only some edits, so work on a copy. Project content such as names, transcripts, markers and frames is sent to the AI service you use. Premiere's generative tools spend Adobe credits, and the agent asks before starting one. Raw ExtendScript is off unless enabled. The mask and speed-keyframe operations bring Premiere to the front and use real clicks, needing Accessibility and Screen Recording permissions; the connector is unsigned, so CEP…

Installation

In SourceWeft

  1. Open Premiere-MCP-IMT + SKILLS — Premiere Pro for AI agents in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

[Premiere-MCP-IMT + SKILLS — Premiere Pro for AI agents]

Premiere-MCP-IMT + SKILLS — Premiere Pro for AI agents

Let an AI assistant (Claude Desktop, Cursor, ChatGPT, Google Antigravity or any MCP client) work inside your running Adobe Premiere Pro: read the project and its sequences, cut, trim, move and ripple clips, apply effects, transitions and keyframes, add markers and captions, reframe, check frames and export — from plain language.

Every Adobe UI language · vision: masks and keyframes by sight · a large skills library · built for generative features. Premiere-MCP-IMT works with every UI language Adobe ships, and the agent sees the frame: it looks at exported frames, draws Object Masks on the people or parts it sees — a click for a whole person, a lasso for a face; add, subtract, track (macOS) — sets speed keyframes (Time Remapping) by dragging them into place on the timeline, and checks keyframed motion frame by frame. Premiere's scripting has no API for masks or speed keyframes; none of the other Premiere MCP servers we compared creates Object Masks, and Adobe's own AI Assistant in Premiere (beta) documents neither. It comes with a large skills library and is built for Adobe's generative features — Generative Extend, Generative Media, Object Mask, Enhance Speech, transcripts and captions, Auto Reframe, Search, the AI Assistant.

Two parts, made to work together:

  • The MCP server — about 440 Premiere operations behind nine tools; writes are verified by reading Premiere back; works on every Premiere UI language (verified live); Object Masks by sight on macOS (the masks operations drive Premiere's own mask tool — there is no API for it); no telemetry.
  • The skill (skills/pr, also served by the pr_skill tool, so it works in any client) — how to drive Premiere safely, where scripting ends, and how to use Premiere's own AI tools: Generative Extend, the Generative Media Tool (video, sound effects, music), Object Mask, Speech to Text and Paper Edit, captions, Enhance Speech and stem separation, Search, the AI Assistant, Auto Reframe, Scene Edit Detection, Morph Cut — plus recipes that combine them and the gotchas.

Built for localized Premiere, not just English. Most Premiere MCP servers and agent skills assume the English UI — on a German, Chinese or Russian install their effect, transition and parameter names stop matching. Premiere-MCP-IMT finds components by matchName and translates English effect, transition and parameter names through Premiere's own UI dictionaries, so the same agent and the same skill work on all ten UI languages Premiere ships — verified live.

By Immersive Media Technologies.

macOS and Windows · Premiere Pro 2024–2026 — verified live on macOS (26.5) and on Windows 11 ARM (26.5.2, Claude Desktop from the Microsoft Store) · Node.js 20.19+ (none needed with the Claude Desktop extension).

[!CAUTION] This edits real Premiere projects, and scripted undo covers only some edits — work on a copy. Project content (names, transcripts, markers, frames) goes to the AI service you use. Premiere's generative tools spend your Adobe credits; the agent asks before starting one. Raw ExtendScript is off unless you turn it on.

What you can ask for

  • "List every clip shorter than 12 frames and every gap in the active sequence."
  • "Tighten the pauses in this interview to a third of a second and put a Morph Cut on the jump cuts where the camera is locked off." (skill)
  • "Make 9:16 and 1:1 versions of this edit and check the titles stay out of the TikTok UI zones."
  • "Blur the face of the man on the left for the whole shot." (macOS: the agent outlines the face with Object Mask itself, tracks it and checks it on other frames; on Windows you click the subject)
  • "Mask everyone in this shot, faces first, and keep the mask on them to the end of the clip."
  • "The music in this interview was recorded on location — get rid of it and score it again under the dialogue." (skill)
  • "Mark every beat of this track and lay the selects to it, two beats per shot."

The agent combines the operations itself and checks the sequence between steps; the skill tells it which Premiere AI tool fits the job, what it costs and where it breaks.

Install

You need Premiere Pro running on this computer and the connector — a small CEP extension inside Premiere that the server talks to. The agent installs it for you on first use (pr_install_connector, it asks first; then quit and reopen Premiere), or the installer below does.

ClientHow
[Image] Claude Desktopdownload premiere-mcp-imt.mcpb, double-click, Install (Windows: drag it onto Settings → Extensions). No Node needed.
[Image] CursorAdd to Cursor — one click: Cursor opens and writes the config (or put the snippet below into ~/.cursor/mcp.json).
[Image] Google AntigravitySettings → Customizations → Installed MCP Servers → View raw config → add the snippet below.
[Image] ChatGPT (remote servers only)on the Premiere machine: npx -p @immersive-media-technologies/premiere-mcp-imt premiere-mcp-imt-chatgpt — it starts the server behind a tunnel and prints the URL to paste into ChatGPT.
Any MCP clientthe snippet below in its mcpServers.
json
{  "mcpServers": {    "premiere": {      "command": "npx",      "args": ["-y", "@immersive-media-technologies/premiere-mcp-imt"]    }  }}

From source, with a live self-test (the route we use):

bash
git clone https://github.com/Immersive-Media-Technologies/premiere-mcp-imt.gitcd premiere-mcp-imt && ./install.sh        # Windows: .\install.ps1

The installer checks Premiere, Node and FFmpeg, builds, installs the connector (asks first), asks the running Premiere whether the connector answers, and — after asking — writes the entry into Claude Desktop's, Cursor's and Antigravity's configs, backing them up.

If something does not connect

  • pr_verify says the connector does not answer: quit Premiere completely after the connector was installed and open it again; close any open Premiere dialog (a modal dialog blocks scripting). premiere-mcp-imt --diagnose-cep checks the files, the debug flags and Premiere's signature log.
  • The connector is unsigned: CEP must allow unsigned extensions (PlayerDebugMode = 1 for CSXS 9–14; on Windows a String (REG_SZ) value under HKCU\Software\Adobe\CSXS.<n>). The installers set it.
  • spawn npx ENOENT on macOS: GUI apps do not see your shell PATH — put the absolute path of a Node from Homebrew or nodejs.org in place of npx (nvm's Node is invisible to them).
  • Loudness, silence (detect_silence), beat and scene-change analysis need FFmpeg (brew install ffmpeg, winget install Gyan.FFmpeg).
  • Windows, Claude Desktop from the Microsoft Store: the Store app's registry writes are sandboxed. If Premiere does not load the connector after pr_install_connector, run powershell -ExecutionPolicy Bypass -File scripts\install-cep.ps1 once from a normal PowerShell in a source checkout (or use install.ps1).
  • premiere-mcp-imt --doctor checks the local install without reading a project.

How it stays safe

Verified writesedits read Premiere back and report verified only when the readback matches; anything else is reported as unconfirmed. An edit whose outcome is unknown pauses further edits until the agent has looked at the project and acknowledged it — a retry never applies a change twice.
Undo where Premiere records it, checkpoints elsewhereEdits that Premiere records in its history (effects added or removed, cuts, inserts) report undoSteps and can be reverted with undo — each step is checked against Premiere's undo stack (experimental QE). Property and marker writes are not recorded: create_sequence_checkpoint clones the sequence before a risky edit; diff_sequence_snapshots shows exactly what changed (added, removed, moved, trimmed clips with frame deltas).
Plan → preview → applyplan_* operations only compute; compound edits apply only with the confirmation token of their preview, unchanged.
Read-only modePR_MCP_READONLY=true — inspection and exports only; the project cannot change.
No raw scripting by defaultpr_execute_script exists only with PR_MCP_ENABLE_SCRIPT=true.
No telemetrythe server sends nothing anywhere.
Real clicks only where there is no APIthe masks operations and time_remap_ui (macOS) bring Premiere to the front and click its mask tool or drag its speed keyframes; they need Accessibility and Screen Recording for the app that runs the server, refuse on other systems, are off in read-only mode (only the overlay inspection stays), and report committed_unverified with an overlay image to check.

How the agent sees

Adobe gives outside agents no picture: its scripting returns a project's structure, not its pixels, and panels have no live access to the Program Monitor (requested since November 2025). Its own connector for AI assistants «operates at Adobe Express level, not Photoshop or Premiere» (MindStudio), and editors who drive Premiere through AI report «no effects or anything more advanced» (Adobe forum).

Premiere-MCP-IMT gives the agent two kinds of sight. The model's: pr_export_frame returns any frame of the sequence as an image, and the model decides from the picture what to mask, where a speed change belongs, whether a keyframed move lands. The server's: Premiere's scripting has no API for masks or speed keyframes — they exist only in its UI. So the server looks at Premiere itself: it captures the window, finds the Program Monitor, the Effect Controls rows and the timeline by their shapes (icons and layout, not labels, so it works on every UI language), acts with real mouse events, and captures again to confirm every step.

Object Masks by sight (macOS): one click of the Object Mask tool on the Program Monitor takes a whole person; a closed lasso takes only the part it outlines (a face, a sign). ⊕ / ⊖ add to or subtract from the mask — the server selects the mask by its row first, because Premiere drops the selection after tracking or a playhead move. Then tracking, Feather, Expansion and the other mask settings; after each edit the server measures the mask's red overlay against the plain exported frame, and the model checks the result frame by frame.

Speed keyframes by sight (Time Remapping, macOS): the server adds a speed keyframe at the playhead in Effect Controls, zooms the timeline until the clip is wide enough to place a frame, finds the keyframe's handle on the clip and Shift-drags it until the measured position matches the target frame — then the model checks the motion on exported frames.

Tools

ToolPurpose
pr_verifyconnector, open project and active sequence — the first call
pr_get_sequencetracks and clips with their node_id, times and speed
pr_catalog / pr_do / pr_batchdiscover the ~440 operations (45 categories) and run one or several
pr_export_framea frame of the sequence as PNG — the agent's eye
pr_save_projectsave
pr_install_connectorcheck / install the connector inside Premiere
pr_skillthe agent skill: the router and its references
pr_execute_scriptraw ExtendScript — only when enabled

PR_MCP_FACADE=0 lists every operation as its own tool instead (more than 600K tokens of definitions — beyond most context windows; the nine tools cost about 1.5K).

VariableDefaultMeaning
PR_MCP_READONLYofftrue → inspection and exports only
PR_MCP_ENABLE_SCRIPTofftrue → pr_execute_script (arbitrary ExtendScript)
PR_MCP_FACADEon0 → every operation as a separate tool
PREMIERE_MCP_CAPABILITIES—explicit profile, e.g. inspect,edit,export,filesystem; wins over the switches
PREMIERE_TEMP_DIROS tempthe connector's file mailbox
PREMIERE_UXP_TOKEN—enables the optional UXP bridge (with the UXP plugin in Premiere) and its *_uxp operations

The skill

skills/pr/SKILL.md is a short router the agent reads first (pr_skill with no arguments); it points to the one reference the task needs (pr_skill {doc}):

SKILL.mdthe tools, the order of work (state → checkpoint → operate → verify by values, then by frame), addressing (node ids, source vs sequence time, selection-driven operations, localized Premiere) and a table of what scripts cannot do in Premiere with the reference for each.
references/generative.mdGenerative Extend, the Generative Media Tool (video with first/last/reference frames, sound effects with a voice guide, soundscape, music), Firefly Boards → sequence, partner models, credits per second.
references/masking.mdObject Mask, why a new mask "does nothing", Clip vs Frame keyframes, Sharpness / Expansion, where mask data lives and how it goes missing.
references/speech-text.mdSpeech to Text, text-based editing of pauses and fillers, Paper Edit, single-word captions, and the server's word-timeline planners (fillers, pauses, retakes, bleeps, captions).
references/audio.mdEnhance Speech, stem separation (beta), ducking by UI and by keyframes, Remix, FFmpeg analysis.
references/timeline-automation.mdAuto Reframe, Scene Edit Detection, Morph Cut (and why it exports as a dissolve), Color Match, Warp Stabilizer.
references/search-assistant.mdthe Search panel (Media Intelligence, similar frames, sound search) and the AI Assistant beta next to an agent.
references/layer-conform.mdsyncing an AI-generated layer with the plate before compositing: time drift (speed keyframes from measured coincidence points), geometry (Motion per piece), colour (Levels), the people mask and seam — measured on exported frames by scripts/conform.py.
references/combos.mdeleven recipes that combine tools: a generated shot between two clips, interview cleanup, a tracked face blur, storyboard to sequence, removing location music, social versions, silence and beat cutting.
references/gotchas.mdavailability by region and licence, credits, beta vs release, handoff to colour and sound, export, scripting limits that look like bugs.

Clients with their own skill support can load the folder directly (e.g. copy skills/pr into ~/.claude/skills/).

Why this one

Several Premiere MCP servers exist; they differ in what the model sees, what it knows and what it sends home. As of 2026-11-04 (the vision row re-checked 2026-11-06):

Premiere-MCP-IMTpremiere-pro-mcp (upstream)AdobePremiereProMCPAdobe_Premiere_Pro_MCPflue
Tools the model sees by default9, over ~440 operationsthe whole catalog (~440 tools)74 (profile standard) of ~1,0605, over 283 operationsnone — the agent pipes ExtendScript
Knowledge of Premiere's AI toolsskill with 8 references + recipes2 skills (editing, development)7 workflow skills1 skill1 skill (scripting)
Non-English Premiereevery UI language (all ten), verified liveEnglish and es-ES measurednot statednot statednot stated
Vision: acting on what the agent seesframes returned as images; Object Masks drawn on what it sees (click = whole person, lasso = a part), ⊕/⊖, tracking, overlay check; speed keyframes placed on the timeline by sight (macOS); keyframes checked frame by frameframes returned as images; Object Mask audits (read-only); no speed keyframes (no scripting setter)frames as images; no masks; a time-remap keyframe tool listed in its unaudited full catalogstill export; no masks; constant speed onlyno (scripting only)
TelemetrynonePostHog, off unless a key is setnone foundon by default (opt-out)none mentioned
Connector installby the agent, with consent; installers with a live self-testCLI flagbuild from source (Go, Rust, Python)CLI flagpip + setup
ClientsClaude Desktop (.mcpb), Cursor, Antigravity, ChatGPT gateway, any MCP clientnpm, .mcpb, Codex / Claude Code pluginsbuild from sourcenpm, plugins, .mcpbCLI + skill
LicenseApache-2.0MITMITMITMIT

Adobe's own AI Assistant in Premiere (public beta since June 2026) organizes projects and assembles rough cuts; its documentation mentions neither Object Masks nor speed keyframes (checked 2026-11-06).

What we changed in the upstream code, besides the skill and the packaging: component and property lookup by matchName and English names translated through Premiere's own UI dictionaries (on a localized Premiere every opacity, motion, keyframe and volume operation of upstream failed), split_clip that cuts the clip you name (upstream cut track 0), no telemetry, stdio only, and a compact tool surface.

Releases

Each version is described on the Releases page; every change is in CHANGELOG.md.

  • v1.3.1 · 2026-11-06 — descriptions name speed keyframes placed by sight next to Object Masks; a «How the agent sees» section.
  • v1.3.0 · 2026-11-06 — layer conform: syncing an AI-generated clip with the original plate — cut-only time (pieces slipped by whole frames, conform.py 1.2.0 with refine-cuts), Transform per piece, Levels on the lower clip, the people mask with its edge inside the silhouette; effect-property tools accept property_index alone; Object Mask edit and Time Remap UI fixes.
  • v1.2.0 · 2026-11-05 — vision: Object Masks by sight on macOS — the agent clicks a whole person or outlines a part (a face) on an exported frame, adds and subtracts, tracks and checks the red overlay; Premiere has no API for masks.
  • v1.1.1 · 2026-10-28 — Premiere's home screen: pr_verify says when no project is open, and an edit tried without a project no longer pauses the edits after it.
  • v1.1.0 · 2026-10-28 — built on premiere-pro-mcp 1.26.0: every edit read back from Premiere, about 60 more operations, the UI-language layer carried into the new code; reinstall the connector with pr_install_connector.
  • v1.0.2 · 2026-10-28 — descriptions everywhere: every Adobe UI language, a large skills library, built for Adobe's generative features.
  • v1.0.1 · 2026-10-28 — product name Premiere-MCP-IMT (was PR-MCP-IMT); the MCP Registry entry also lists the npm package; the package is on npm from 1.0.0.
  • v1.0.0 · 2026-10-28 — every Premiere UI language (English names translated through Premiere's own dictionaries; verified live), honest results for unknown effects, colour correction and LUTs, Apache-2.0.
  • v0.1.0 · 2026-10-26 — first public release: the Premiere agent skill, nine-tool surface, pr_skill, connector install by the agent, Claude Desktop extension, installers with a live self-test, ChatGPT gateway.

npm test runs offline — no Premiere needed.

License

Apache License 2.0 — © Immersive Media Technologies.

  • Free to use, modify and distribute, commercial use included.
  • Attribution is required. Every redistribution of this software or of a work derived from it must carry the notice from NOTICE: Premiere-MCP-IMT by Immersive Media Technologies — https://github.com/Immersive-Media-Technologies.
  • No warranty.

Third-party notices (upstream MIT code): THIRD-PARTY-NOTICES.md. Adobe, Premiere Pro and Firefly are trademarks of Adobe Inc.; this project is not affiliated with or endorsed by Adobe.

Source: README.md at commit ce661b3

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

3
  1. v1.3.1LatestNov 7, 2026
  2. v1.2.0Nov 5, 2026
  3. v1.1.1Oct 29, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts