ftx-mcp

Talk to your FactoryTalk Optix project. ftx-mcp connects AI tools (Desktop, Cowork, Code — or any MCP client) to FactoryTalk Optix Studio on your Windows machine, so you can build and change HMI screens by describing what you want:

"Add a header that says 'Hello Optix' to Screen 1 and show me."

"Bind that label's visibility to Model/PumpRunning."

"Looks right. Launch the emulator and validate it."

Your LLM authors the change directly into your open Studio project, runs the emulator, and looks at the rendered runtime to confirm it worked. It is a development and testing companion: shipping to hardware stays in Studio's own Deploy dialog, in your hands. Everything besides the LLM calls runs locally on your machine. No cloud service, no API key; your LLM of choice provides the intelligence.

Blog Post

Install (10 minutes)

Requirements: Windows 11, FactoryTalk Optix Studio 1.7.x, and an MCP client like Claude Cowork. Also: Google Chrome (the CDP verify loop; skip with setup.ps1 -NoCdp) and — for the Claude Desktop Microsoft Store build connector path — Node.js (winget install OpenJS.NodeJS.LTS; the config uses npx mcp-remote). Setup auto-installs Tesseract OCR via winget for the zero-vision-token text tools (skip with -NoOcr; everything else works without it) and Pillow into the venv for pixel diffing.

Run setup.ps1 from a regular PowerShell window — not a shell hosted inside a packaged app like the Store build of Claude Desktop, whose %LOCALAPPDATA% writes are virtualized (setup now detects this and refuses). If you downloaded a ZIP instead of cloning, run Get-ChildItem -Recurse | Unblock-File first.

If setup.ps1 (or any bootstrap\*.ps1) fails to run at all with a script- execution / "running scripts is disabled" error, your machine's execution policy is blocking unsigned scripts. Unblock them for the current window only — no admin, no persistent policy change — then re-run:

Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass

git clone https://github.com/asqi-carter/ftx-mcp.git
cd ftx-mcp
.\bootstrap\setup.ps1                                  # install (does not start anything)
.\bootstrap\services.ps1 start                         # start the service + cdp chrome
.\bootstrap\services.ps1 status                        # verify health
http://127.0.0.1:8765/ui                               # health dashboard

Auth: the default install is loopback-only with auth OFF — setup never prompts, and no token is needed on your own box. Binding to the LAN is the one case that requires auth: install with .\bootstrap\setup.ps1 -EnableAuth (see docs/security.md). If a UI/API unexpectedly asks for a bearer token, a previous install enabled auth — re-run setup with -NoAuth to clear it.

To remove an install (or reset before a clean reinstall): .\bootstrap\uninstall.ps1 stops and unregisters the scheduled tasks and reaps any leftover CDP chrome; add -All to also delete state (issued tokens, chrome profile, persisted auth choice) and the venv.

Start the Studio Bridge (5 minutes)

1. One-time bridge setup (per project): in the Studio project tree, right-click NetLogic → add a new DesignTime NetLogic named StudioMCPBridge, double-click it to open the C# editor, and paste in studio-bridge/StudioMCPBridge.cs (make sure to rebuild in your code editor or save in Studio to trigger a rebuild)
2. Setup the Project once per Studio session: right-click StudioMCPBridge → SetupProject → This just adds a webui for validation access at localhost:8081
3. Start the bridge once per Studio session: right-click StudioMCPBridge → StartBridge → accept Studio's one-time security prompt.
4. Verify bridge health: Studio Output will show listening on http://127.0.0.1:8768 (the bridge). The service dashboard at http://127.0.0.1:8765/ui shows bridge status too.

Cowork (5 minutes)

Requirements: MCP server and bridge running. Claude desktop app downloaded.

cd ftx-mcp                   # install + start the service
.\bootstrap\setup-mcp-client.ps1 -WriteConfig          # adds as a connector to desktop app

Restart Claude Desktop (You might have to end task in task manager to fully restart), then ask Claude to "run optix_status(action='doctor')" — it reports anything missing, with a plain-English fix for each item. In settings > connectors you can adjust the permissions for each of the tools

Claude Code (5 minutes)

Requirements: MCP server healthy and claude accessible in cli

claude mcp add --transport http ftx-mcp http://127.0.0.1:8766/mcp

Then in Claude Code, run /mcp and confirm ftx-mcp shows as connected.

Visual Studio Code (5 minutes)

Create or open .vscode/mcp.json and add:

{
  "servers": {
    "ftx-mcp": {
      "type": "http",
      "url": "http://127.0.0.1:8766/mcp"
    }
  }
}

Your first change

1. Ask for a change, e.g. "Using the ftx mcp, Add a Start and stop button that toggles Model/MotorRun on MainWindow, and verify it works with a label with the text of 'MOTOR RUNNING' that has visibility tied to MotorRun."
2. Watch it work: author → emulator preview → screenshot → and when it looks right, you deploy it to your hardware from Studio as usual.

What it's capable of

• Author widgets, properties, bindings, computed expressions, events, translations, and multi-screen navigation — live in the open Studio project.
• Run Studio's built-in emulator via F5 key and read the runtime log.
• Verify by looking at the webui: screenshot, click buttons and tabs, type into fields.
• Hand back to you to ship when the preview looks right, you deploy from Studio as usual. This distribution only runs the emulator.

The full tool list (28 tools by default — gate env vars can add more — plus the same surface over plain HTTP for scripts and CI) is in docs/tool-reference.md. Token economy: screenshots cost ~1-2k vision tokens each; the OCR tool family (optix_cdp_read_text, find_text, navigate, sweep/diff) turns most checks into free text reads — see the optix-blind-authoring and optix-visual-regression playbooks for the workflow (setup installs tesseract automatically; -NoOcr opts out).

Bundled authoring playbooks (navigation, bound controls, styles, expressions) ship with the server itself — Claude discovers them via optix_list_skills / optix_get_skill, so every connected client gets them with zero setup. (In Claude Code they also load natively as skills.)

Security & safety posture

• Local only. The service binds 127.0.0.1 and talks to nothing off the machine. Optional bearer-token auth, enforced before any LAN bind.
• Read-only by default. Every tool carries MCP readOnlyHint / destructiveHint annotations (contract-pinned by tests) so your MCP host can auto-approve reads and gate writes.
• Write gates, not hope. Undeclared properties, array writes, duplicate names, and unsafe re-parents are refused with typed errors; composite operations roll back on failure. File-level edits are refused while Studio has the project open.
• Audited. Every model mutation appends a JSON line to a local audit trail (%LOCALAPPDATA%\ftx-mcp\logs\audit.jsonl) what, when, outcome.
• Shipping stays in your hands. Previewing never touches your runtime; deploying to hardware happens from Studio, full stop. This distribution does design time edits and testing via the emulator.

The full posture including the prompt-injection surface analysis is in SECURITY.md.

Documentation

Compatibility

Tested with FactoryTalk Optix Studio 1.7.x on Windows 11, Python 3.12. Optix CLI behavior is not contract-stable across major versions — pin your Studio version in production.

License

MIT · © 2026 ASQI · Not affiliated with or endorsed by Rockwell Automation. FactoryTalk Optix is a trademark of Rockwell Automation, Inc.; this project orchestrates locally installed Optix binaries without redistributing them. See NOTICE.