Adobe MCP

adobe-mcp — Adobe Creative Cloud MCP server (macOS)

Drives After Effects, Photoshop, Illustrator, Premiere Pro (and Character Animator puppet authoring) from Claude, plus Blender and Substance 3D Painter for 3D look-dev and cinematic animation. See ROADMAP.md for the full tool matrix and what is next. Bridge: MCP client → this server → AppleScript DoScript → AE → temp result file → back.

Tools

3D & cinematic VFX pipeline (Blender + Substance 3D Painter)

Two Python bridges extend the server beyond ExtendScript so one agent can take a shot from blocking to final composite:

                  ┌──► Blender ──────────► Python (bpy): live add-on socket :9877, or headless `blender -b`

MCP client ───────┼──► Substance Painter ─► Python: remote scripting HTTP :60041 (/run.json) └──► After Effects ────► ExtendScript via osascript (as above)

Free CC0 texture & HDRI libraries (Poly Haven, ambientCG): everything both publish is public domain, so it's free for commercial work with no attribution.

The downloaded folder also works as texture_dir for Creature_Walk_Cinematic.py. Set POLYHAVEN_API / AMBIENTCG_API to point at a mirror if you need to.

Library scripts (run with run_script, parameters via params):

• scripts/blender/Creature_Walk_Cinematic.py: a heavy-biped walk timed from physics. Speed comes from the Froude number, stride from Alexander (1976), and feet stay planted with zero slide. The neck and tail lag 3 frames per joint, the head is gaze-stabilised and turns toward the lens, and the chest breathes. The camera is a handheld rig with operator lag and footfall impact shake. The render setup is Cycles with a 180° shutter, a shadow catcher, and beauty/shadow/mist PNG sequences plus a multilayer EXR. Tested headless against Blender 5.0.1 (bpy); it also handles the 4.x APIs.
• scripts/substance/Creature_Skin_LookDev.py: optional bake, then a matte "aged scales" base, a cavity-masked wetness layer, and glossy eye/mouth sets, all in one undo step. Optional export.
• scripts/ae/Cinematic_3D_Composite.jsx: imports the passes, multiplies the shadow, adds a precomped edge light wrap, depth haze from mist, an optional depth lens blur, CG softening, and grain on the CG only.

Setup:

• Blender live mode: Edit → Preferences → Add-ons → Install from Disk → bridges/blender/adobe_mcp_bridge.py, then enable it (listens on 127.0.0.1:9877; override with BLENDER_MCP_PORT). Background mode needs no add-on; it finds /Applications/Blender*.app or blender on PATH (override with BLENDER_PATH).
• Substance 3D Painter: launch with remote scripting enabled: open -a "Adobe Substance 3D Painter" --args --enable-remote-scripting (port override: SUBSTANCE_PAINTER_PORT).
• Both bridges are local-only and unauthenticated: anything on the machine that can reach the port can run Python in the app. Disable the Blender add-on or quit Painter when you're done.

Knowledge system (the server learns as it goes)

The server ships with a curated knowledge base (knowledge/lessons.jsonl) seeded from real production sessions — matchName gotchas, API quirks, parameter-type traps, and techniques (blueprint line extraction, seamless wind loops, QE transitions, Auto Reframe...).

Three mechanisms keep it useful and growing:

1. Auto-surfaced gotchas — every failed *_run_extendscript call is matched against the lesson keywords; hits are appended to the error message the client sees (e.g. hit Illegal Parameter type in Premiere and the fix for autoReframeSequence's string preset arrives with the error). Failures are also logged to $WORKSPACE/error-log.jsonl as raw material for new lessons.
2. `knowledge_search` / `knowledge_add` tools — agents consult lessons before attempting unusual scripting, and record new gotchas when they solve one that wasn't auto-suggested. Runtime lessons persist to $WORKSPACE/learned.jsonl (survives repo updates, per-install).
3. Server instructions + tool descriptions — the highest-value rules (explicit document targeting, visual verification, render-timeout handling, ES3 limits) are baked into the MCP instructions and the per-tool descriptions, so a fresh install starts smart.

The scripts/ per-app libraries are the executable side of the same idea — distilled, reusable skills (list_scripts / run_script): blueprint line extraction (PS), seamless wind loops and kinetic typography rigs (AE), one-call vertical Auto Reframe (Premiere).

Install

Option A — one-click for Claude Desktop & Cowork (no Node needed)

1. Have the Adobe apps you want to drive installed. In After Effects, enable Preferences → Scripting & Expressions → "Allow Scripts to Write Files and Access Network".
2. Download the latest adobe-mcp-x.y.z.mcpb from Releases and double-click it (or drag it onto Claude Desktop → Settings → Extensions). Claude Desktop bundles its own Node runtime — nothing else to install.
3. Use it. The first tool call for each app triggers a one-time macOS Automation permission prompt — click Allow.

Upgrading? Remove the old version first (Settings → Extensions → Adobe Creative Cloud → Uninstall), then install the new .mcpb — in-place installs over an existing version can leave stale files behind while reporting the new version number.

Option B — any MCP client via npm

Requires Node 18+ installed system-wide (nodejs.org installer or brew install node — note GUI apps often cannot see nvm-managed installs). No clone, no npm install: every client config below just runs

npx -y @ajsmonty/adobe-mcp

Option C — from source (contributors)

git clone https://github.com/AJSMonty/adobe-mcp.git ~/tools/adobe-mcp

cd ~/tools/adobe-mcp && npm install

point your client at: node ~/tools/adobe-mcp/server.mjs

Premiere Pro only: the ppro_* tools need the bundled CEP panel installed once — copy cep/mcp-bridge to ~/Library/Application Support/Adobe/CEP/extensions/ and enable PlayerDebugMode (see Notes), then open Window → Extensions → MCP Bridge inside Premiere.

Hook up to your MCP client

This is a standard stdio MCP server — any MCP-capable client can drive it (Claude Code, Claude Desktop, Cursor, Windsurf, VS Code Copilot agent mode, Zed, Cline, Goose, ...). Everything below is the same one-liner expressed in each client's config format: run npx -y @ajsmonty/adobe-mcp over stdio (or node .../server.mjs for a source checkout).

Claude Code

claude mcp add --scope user adobe -- npx -y @ajsmonty/adobe-mcp

--scope user = available in all projects; use --scope project from inside a repo to share via .mcp.json with your team. Verify: claude mcp list.

Claude Desktop

Settings → Developer → Edit Config, add to claude_desktop_config.json:

{

"mcpServers": { "adobe": { "command": "npx", "args": ["-y", "@ajsmonty/adobe-mcp"] } } }

Fully quit and reopen Claude Desktop.

Cursor

Settings → MCP → Add server, or edit ~/.cursor/mcp.json (global) / .cursor/mcp.json (per-project):

{

"mcpServers": { "adobe": { "command": "npx", "args": ["-y", "@ajsmonty/adobe-mcp"] } } }

Enable it under Settings → MCP, then use Agent mode (composer) — tool calls appear with approval prompts.

VS Code (Copilot agent mode)

.vscode/mcp.json in the workspace (or user mcp.json via the "MCP: Add Server" command):

{

"servers": { "adobe": { "type": "stdio", "command": "npx", "args": ["-y", "@ajsmonty/adobe-mcp"] } } }

Windsurf

~/.codeium/windsurf/mcp_config.json — same mcpServers shape as Claude Desktop/Cursor above.

Anything else

If the client supports MCP over stdio, point it at npx -y @ajsmonty/adobe-mcp and you are done.

Client caveats

• macOS Automation permission is per host app. The first tool call from each new client (Cursor, VS Code, a different terminal) triggers its own Automation prompt for controlling the Adobe apps — approve it once per host in System Settings → Privacy & Security → Automation.
• Inline image previews (ae_save_frame, ps_save_preview, ai_save_preview, ppro_save_frame) return MCP image content. Claude clients and Cursor render these inline; clients without image rendering still get the metadata text, and the model can fall back to exporting a file and reading it from disk.
• Server instructions (the working rules + knowledge-base pointers) are sent via the MCP instructions field; most clients pass them to the model, but if yours does not, the same rules are baked into the tool descriptions and knowledge_search.

Script library ("skills")

Drop .jsx / .jsxbin files into scripts/ — they're immediately available via list_ae_scripts / run_ae_script.

• Action scripts (operate on the current selection, no UI) run headlessly — set up selection first via run_extendscript.
• ScriptUI panel scripts open their palette in the AE window for a human to click.
• Plain .jsx panels with a separable core function can often be driven headlessly by evaluating the source with the UI kickoff stripped.

⚠️ scripts/ may contain third-party/purchased tools. Keep this repo private unless you've cleared redistribution rights, or add scripts/ to .gitignore.

Notes

• AE app auto-detected from `/Ap