This is the manual of Studio 1.1.0, an earlier version. The manual of the latest version

The AI assistant

An AI can work on the twin with the same tools you have: it sets up, writes and checks programs and runs them in the simulation. You can talk to one in Studio's own assistant, or connect an AI client such as Claude Desktop. Anything that would change the window's machine waits for your yes. The AI is not part of any safety function.

Turning the AI on

The AI is off in a new installation. To turn it on, do one of these:

click the AI's state in the top bar (AI off) and choose The window's AI switch;
Tools > Settings… > AI and MCP: AI switch, then OK;
send a message in the assistant: this turns the switch on.

Studio keeps the switch as you leave it for the next start. The AI menu (AI assistant…, AI activity…, Connect AI clients…) shows from the Operator level on.

With the switch off, AI clients can still use practice machines in the background, but not the machine in your window.

The AI service and its key

The assistant talks to an AI service you choose, with your own account:

Anthropic (Claude), OpenAI, DeepSeek, Qwen (Tongyi), Kimi (Moonshot AI), Zhipu (GLM);
LM Studio (local) or Ollama (local): a model on your own computer, no key needed;
Custom (OpenAI-compatible): any service with an OpenAI-compatible interface, at the address you type.

To set it up:

Open AI > AI assistant…
Choose the Provider. Studio fills in its Base URL.
Paste your key in API key and click Save key. Result: the key is stored on this computer, encrypted. On Windows only your Windows account can read it. On Linux it goes into the desktop's keyring (Secret Service: GNOME Keyring, KWallet); without one it is encrypted with a key in a file only your user account can read, and the line under the key says so.
Click List models and choose a Model.

Studio does not write the key into its settings, the registry or the audit log. To delete a saved key, empty the field and click Save key.

Note

What you send the AI service goes to that service under your account and its terms: your messages, and with Attach the window's program, setup and check results on, the program, the setup and the check results, and the results of the tools the AI uses. To keep everything on your computer, use a local model.

Before your first message to a cloud service, the assistant says once what goes there. Click Send to go on, or Use a local model… to switch to LM Studio; your message stays in the box. Tools > Settings… > AI and MCP lists the services you agreed to; Ask again makes the assistant ask again.

The assistant panel

AI > AI assistant… opens the assistant in a window of its own.

Type in the box at the bottom. Enter sends; Shift+Enter starts a new line. Stop stops the answer.
Attach the window's program, setup and check results (when they changed) gives the AI what is in your window with each message. It is on by default.
The conversation shows each tool the AI uses, with ✓ when it worked and ✗ when it did not.
Pictures the AI makes (the workpiece, the toolpath, the deviation from the part, a comparison) appear in the conversation. Double-click one to enlarge it or save it.
New conversation starts afresh. Conversations are not saved when you close Studio.

Closing the window only hides it; the conversation stays until you quit.

What the AI can and cannot do

With the switch on, the AI can:

read the machines' state, and home, unlock, jog and send lines of G-code;
set up the blank, the tools, the work zero and the part, probe and inspect;
check, simulate, optimize, generate, edit and load programs, and start or restart a run;
read runs and their results, compare them and make pictures of them;
read the audit log and the licence's state.

Tools that belong to a paid edition (inspection, reports, nesting, the optimizer) work only in that edition.

The AI cannot change Studio's settings, the language, the units, the user level, the AI switch, how Studio asks you, the licence or the audit log.

Asking before it changes the machine

Before the AI changes the window's machine (in the AI's menu in the top bar and in Tools > Settings… > AI and MCP) has two choices:

Ask me every time (the default);
Do not ask (the AI works the window directly).

When the AI wants to change the window's machine (move it, power it on, release the E-stop, load or run a program, change the setup or the tools), a bar opens over the machine bar. It says which AI asks and what it wants to do, and counts down.

Read the question.
Click Allow or Refuse. Enter does not answer: use the mouse, or Tab to a key and press Space.

No answer within 60 s counts as Refuse. A yes also covers that action's own follow-up requests for two minutes.

Studio never asks before the AI stops, holds or resets the machine, presses the E-stop or only reads. Practice machines and the hidden copies the AI simulates on are not asked about.

Read-only on a real machine

When the window's twin follows a real machine (shadow mode, a preview in 1.0; see A real machine), the AI only reads and analyses. It cannot move the machine, change its setup or work offsets, or run programs there, and it does not ask: Studio refuses those requests. Practice machines work as before.

AI activity

AI > AI activity… shows what AI clients do on this computer:

Background machines: the practice machines the AI uses, which you do not see in the window ("window" is the machine in your window);
Runs: the last 16 simulations and optimizations;
Connected AI clients: who is connected, how, and their last call.

From here you can stop a run, look at a run's picture, close a background machine, and stop or start the twin host. The twin host is the background program that holds the practice machines. It starts when an AI client connects and quits after 30 minutes unused.

The dock's AI activity tab lists the AI's calls on the window's machine: time, call, machine, who and result.

Connecting other AI clients (MCP)

AI clients that speak MCP (Model Context Protocol) can use Studio's tools.

Open AI > Connect AI clients…
Click your client: Claude Desktop, ChatGPT desktop (Codex and Work modes), Cursor, VS Code (GitHub Copilot) or LM Studio (local open models). Result: for Claude Desktop and ChatGPT desktop, Studio shows the change it will make to the client's settings and makes it when you click OK; it keeps the client's old file next to it. The others open their own install page.
Quit the client completely and open it again.
Turn on The window's AI switch in Studio.

For any other client, Other clients: copy the configuration copies the settings to paste into it. The client starts TwinwrightStudio.exe --mcp in the background; The command and configuration shows the exact command.

Clients that take only an address

For clients that take only an address, turn on Local URL (HTTP, for clients that only take a URL) in the AI's menu and use Copy the local URL (HTTP). The address works only on this computer and needs the token that comes with it. New token makes the old one stop working.

The tools, their arguments and their results: MCP tools.

What is recorded

Tools > Audit log… (Ctrl+L) records what the AI does, with the AI's name in the Who column. Filter Who by AI to see only the AI's records; double-click a call to see all its records together.

At the default Standard level, the log keeps:

turning the AI switch on and off, and changing how Studio asks;
each question, your answer and each AI action on the window's machine;
each tool call, with a short form of its arguments;
clients connecting and leaving;
the assistant's messages: the first 180 bytes in the summary, up to 2000 bytes in the record.

Keys, tokens and passwords are written as ***.