illustrator-mcp

mcp
Guvenlik Denetimi
Uyari
Health Uyari
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 6 GitHub stars
Code Uyari
  • process.env — Environment variable access in src/index.ts
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

Control a real, already-open Adobe Illustrator document with natural language via MCP, from Claude, ChatGPT, Cursor, or any client. Not an image generator: it edits actual layers, shapes and text through Illustrator's own scripting engine.

README.md

Illustrator MCP

Give Claude, ChatGPT, or Cursor the keyboard and mouse for a real, already-open Adobe Illustrator document.

This is an MCP (Model Context Protocol) server that bridges an AI assistant and Adobe Illustrator itself, not a generator that hands you a picture and calls it done. Ask your assistant to "create an A4 poster, add a red rounded rectangle and a headline" and it drives your actual, running copy of Illustrator through Illustrator's own scripting engine (ExtendScript). What comes out is the same file you can keep editing by hand afterward, layers and all.

Not just image generation. Plenty of tools, OpenAI's own image models included, turn a prompt into a vector-looking picture. That picture has no relationship to anything you're working on: no layers, no editable text, no artboards, nothing left to select or nudge. Illustrator MCP is a different kind of thing. It opens, reads, and edits a real Illustrator document while it's sitting open on your screen. Vectorizing a raster image is one of its 29 tools, not the whole project.

Works on macOS and Windows. Illustrator must be installed on the same machine as the server.


What can it do?

Say things like:

  • "Create a 1080×1080 RGB document and add the title 'SALE' in bold orange."
  • "Draw three circles in a row, then align them to the center of the artboard."
  • "Open campaign.ai, find the layer called 'Pricing', and change the numbers."
  • "Place ~/logo.png in the top-left corner and scale it to 200px wide."
  • "Select everything named 'badge' and make it blue."
  • "Export the current artboard as a 2x transparent PNG to my Desktop."
  • "Read all the text objects in this document and list their contents."

Every one of those runs against the Illustrator window that's actually open in
front of you, not a sandbox or a throwaway render. The server ships 29 tools
covering documents, artboards, layers, shapes, text, images, transforms,
alignment, color, selection, and export, plus a run_script escape hatch that
can execute any Illustrator ExtendScript, so an agent is never boxed in. Image
vectorization (illustrator_vectorize_image, Image Trace under the hood,
expanded into real editable paths) is one of those 29 tools: handy when a client
hands you a raster logo and you need clean paths, but a feature, not the product.


How it works

┌─────────────┐   MCP (stdio)   ┌──────────────────┐   ExtendScript   ┌──────────────┐
│ Claude /    │ ◄────────────► │ illustrator-mcp   │ ◄─────────────► │  Adobe        │
│ ChatGPT /   │   tool calls    │ server (Node/TS)  │  osascript /     │  Illustrator  │
│ Cursor ...  │                 │                   │  PowerShell COM  │              │
└─────────────┘                 └──────────────────┘                  └──────────────┘
  • The AI client sends tool calls over MCP.
  • The server turns each call into a small ExtendScript program.
  • It runs that script inside Illustrator via AppleScript do javascript (macOS) or COM DoJavaScript (Windows).
  • Results come back as structured JSON.

Because the final layer is ExtendScript, anything the Illustrator scripting API supports is reachable.


Requirements

  • Adobe Illustrator installed (any reasonably recent version) and, ideally, running.
  • Node.js 18+.
  • macOS or Windows.
  • An MCP-capable client (Claude Desktop, Claude Code, Cursor, ChatGPT desktop with MCP, …).

Install

git clone https://github.com/gherardo200-glitch/illustrator-mcp.git
cd illustrator-mcp
npm install        # also builds via the "prepare" script
npm run build      # (only needed if you skip install-time build)

This produces dist/index.js, the server entry point. Note its absolute path: you'll need it for the client config, e.g.:

/absolute/path/to/illustrator-mcp/dist/index.js

Connect it to your AI client

Claude Desktop

Edit the config file:

  • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
  • Windows: %APPDATA%\Claude\claude_desktop_config.json
{
  "mcpServers": {
    "illustrator": {
      "command": "node",
      "args": ["/absolute/path/to/illustrator-mcp/dist/index.js"]
    }
  }
}

Restart Claude Desktop. You should see the Illustrator tools appear.

Claude Code

claude mcp add illustrator -- node /absolute/path/to/illustrator-mcp/dist/index.js

…or commit a .mcp.json to your project (see examples/mcp.json).

Cursor

Add to ~/.cursor/mcp.json (or the project's .cursor/mcp.json):

{
  "mcpServers": {
    "illustrator": {
      "command": "node",
      "args": ["/absolute/path/to/illustrator-mcp/dist/index.js"]
    }
  }
}

ChatGPT (desktop or web)

ChatGPT reaches MCP servers through connectors (Developer Mode, Plus/Pro or
Business/Enterprise/Edu). It does not launch a local stdio server directly.
The clean way to connect this local server is OpenAI's Secure MCP Tunnel
(tunnel-client): outbound-only, no public port. This repo ships an HTTP
transport
for exactly this: run node dist/index.js --http (listens on
http://127.0.0.1:3000/mcp), though the tunnel can also drive the stdio server
directly.

Step-by-step: docs/ATTIVAZIONE.md (Italian) ·
technical notes: docs/CHATGPT.md.

Any other MCP client

Two transports, one codebase:

  • stdio (default): node dist/index.js
  • Streamable HTTP: node dist/index.js --http (PORT / HOST override the
    default 127.0.0.1:3000, endpoint /mcp, /healthz for liveness). Rejects
    requests with an unrecognized Host/Origin header (DNS-rebinding
    protection); set MCP_HTTP_TOKEN to also require an X-MCP-Token header on
    every request. See Safety notes.

First run & permissions (macOS)

The first time the server talks to Illustrator, macOS asks permission to let your
terminal/app control Adobe Illustrator (Apple Events automation). Click OK.

If you denied it or nothing happens, enable it manually:
System Settings → Privacy & Security → Automation → (your app: Terminal / Claude / Cursor) → enable Adobe Illustrator.

Verify the connection by asking your assistant to run illustrator_get_status
(it does not launch Illustrator, so it's a safe first call).


Tools

Area Tools
App illustrator_get_status
Documents illustrator_create_document, illustrator_open_document, illustrator_list_documents, illustrator_save_document, illustrator_close_document, illustrator_export_document
Inspect illustrator_get_document_info, illustrator_get_selection, illustrator_list_layers, illustrator_list_artboards
Create illustrator_create_rectangle, illustrator_create_ellipse, illustrator_create_line, illustrator_create_text, illustrator_place_image
Vectorize illustrator_vectorize_image (Image Trace → editable paths)
Edit illustrator_transform_selection, illustrator_set_color, illustrator_arrange_selection, illustrator_align_selection, illustrator_delete_selection, illustrator_group_selection
Layers illustrator_create_layer, illustrator_set_active_layer
Selection illustrator_select_all, illustrator_deselect_all, illustrator_select_by_name
Power illustrator_run_script

Coordinates: all positions and sizes are in points (= pixels at 72 dpi),
measured from the top-left of the active artboard, with Y increasing downward,
the intuitive convention for design tools. (The run_script helpers convert
from Illustrator's native Y-up global system.)

Colors: hex (#FF7F00), a common name (red, blue, green, …), or none.
Colors are automatically converted to RGB or CMYK to match the document.

Image vectorization (two engines)

illustrator_vectorize_image supports two engines:

  • vtracer (default): the open-source VTracer
    CLI produces a clean SVG that is opened in Illustrator as editable paths. Free,
    local, and usually cleaner than Image Trace. Requires the vtracer binary.
  • image_trace: Illustrator's built-in Image Trace. No install, kept as a fallback.

Install VTracer (once), pick one:

  • Prebuilt binary (no Rust needed): download your platform's archive from the
    releases page, extract the
    vtracer binary, and put it on your PATH (or point VTRACER_PATH at it).
    On macOS you may need xattr -d com.apple.quarantine /path/to/vtracer.
  • From source: cargo install vtracer (needs the Rust toolchain).

The server finds VTracer via PATH or the VTRACER_PATH environment variable. If
it's missing, the tool returns clear install instructions and you can fall back to
engine: "image_trace".

Both engines are best on logos, flat art, line art, and few-color graphics;
photorealistic images produce many paths and usually need manual cleanup.

The run_script escape hatch

illustrator_run_script runs arbitrary ExtendScript. The code runs inside a
function. return a JSON-serializable value to get it back. Prelude helpers are
available: __doc(), __color(hex), __setPos(item, x, y), __itemInfo(item),
__abRect(), __activeAB(doc), __style(item, opts).

// e.g. count every text frame and return their contents
var d = __doc();
var out = [];
for (var i = 0; i < d.textFrames.length; i++) out.push(d.textFrames[i].contents);
return { count: out.length, texts: out };

Safety notes

  • Tools carry MCP annotations (readOnlyHint, destructiveHint, …) so clients
    can prompt before destructive actions (delete, close, run_script).
  • run_script is powerful: it can modify/delete artwork and touch the file
    system (read, write, or delete any file your OS user can access — ExtendScript's
    File/Folder API is not sandboxed). Prefer the specialized tools when they
    fit, and review scripts before approving them, especially from an AI client
    you don't fully trust.
  • In --http mode the server opens a real network port (default
    127.0.0.1:3000/mcp). Anyone who can reach that port can call any tool,
    including run_script. The server now rejects requests whose Host or
    Origin header isn't 127.0.0.1/localhost (plus whatever HOST you set)
    on that port, which is the standard defense against DNS-rebinding
    attacks
    (a page on an attacker-controlled domain that briefly resolves to
    127.0.0.1, hoping to reach your local server through the browser): the
    browser still sends the attacker's real hostname in Host/Origin, so it
    never matches the allowlist and gets a 403. There is still no
    authentication by default
    , so anything else that can reach the port
    with a correct Host header (another local process, or a device on your
    LAN/internet if you port-forward it) can still call every tool. Set
    MCP_HTTP_TOKEN to a random secret to require a matching X-MCP-Token
    header on every request (plain requests get 401); leave it unset and
    behaviour is unchanged. Keep HOST at its 127.0.0.1 default, don't
    port-forward or expose 3000 on your LAN/internet, only run --http mode
    while you're actively using it, and only use OpenAI's outbound-only Secure
    MCP Tunnel to reach it (see docs/CHATGPT.md; note the
    tunnel client doesn't document a way to attach a custom header today, so
    MCP_HTTP_TOKEN and the ChatGPT tunnel may not combine unless you add a
    small local proxy that injects it). stdio mode (the default, used by Claude/
    Cursor) does not open a network port and is not affected.

Troubleshooting

Symptom Fix
"Not authorized to control Illustrator" Grant Automation permission (see above), then retry.
"Could not reach Adobe Illustrator" Make sure Illustrator is installed and running.
"No document is open" Create one with illustrator_create_document or open a file first.
Calls time out Illustrator may be showing a modal dialog; bring it to the front and dismiss it.
Tools don't appear in the client Check the absolute path in the config and restart the client.

Roadmap

  • Chat panel inside Illustrator (UXP/CEP): a dockable panel with a chat box so
    you can talk to the assistant without leaving Illustrator. See
    docs/IN_APP_CHAT.md for the design, integration plan, and
    a cost breakdown (in Italian).
  • More workflow tools (pathfinder, effects, symbols, swatches, batch export).
  • ✅ HTTP (Streamable) transport for ChatGPT / remote MCP clients (--http).

Contributions welcome: see CONTRIBUTING.md.


🇮🇹 Guida rapida (Italiano)

Questo è un connettore che permette a Claude, ChatGPT o Cursor di comandare Adobe
Illustrator
parlandogli in linguaggio naturale ("crea un poster A4, aggiungi un
rettangolo rosso e un titolo"). Non è un generatore di immagini vettoriali: lavora
dentro un documento Illustrator vero, già aperto sul tuo schermo, che resta
modificabile a mano dopo. La vettorizzazione di un'immagine raster è solo uno dei
29 tool disponibili.

Per usarlo:

  1. Installa Node.js 18+ e assicurati di avere Illustrator.
  2. Nel terminale:
    git clone https://github.com/gherardo200-glitch/illustrator-mcp.git
    cd illustrator-mcp
    npm install
    
  3. Copia il percorso assoluto di dist/index.js.
  4. Incollalo nella configurazione del tuo client (vedi Claude Desktop sopra).
  5. Riavvia il client e chiedi: "usa illustrator_get_status" per verificare che
    funzioni. Su macOS, al primo utilizzo concedi il permesso di automazione.

Da lì in poi puoi chiedere di creare documenti, forme, testi, allineare, colorare,
esportare PNG/PDF e altro. Per operazioni avanzate c'è illustrator_run_script,
che esegue qualsiasi script di Illustrator.


License

MIT © 2026

Yorumlar (0)

Sonuc bulunamadi