Autograph MCP Server

MCP (Model Context Protocol) is an open standard that lets an LLM (Large Language Model), also known as an AI app or assistant, connect to an external application and use that application’s tools. The Autograph MCP Server uses this standard to connect your AI app to Autograph, so you can describe a task in plain language and watch the result appear in Autograph.

The AI app (for example, Claude Desktop or ChatGPT Desktop) is called the client. The client talks to the MCP Server through a defined set of tools, such as “create a star shape” or “add a Noise Generator,” and Autograph carries out each tool call on your open project.

The Autograph MCP Server does best with well-defined, laborious tasks, such as:

• Renaming and reorganizing a hierarchy
• Translating the names of assets that arrived in another language

These procedural, often repetitive jobs are the kind you’d happily hand to an assistant. Tasks like X are worth having an AI assistant do in a couple minutes rather than losing an afternoon doing the same work manually. Similarly, routine maintenance should take a minute or two. 

No. You keep working the way you do now, and you decide when to hand something off. The Autograph MCP Server is off by default and when enabled, only takes on specific tasks you describe. The brief, the look, and every creative call remain yours. 

No. The Autograph MCP Server is built into Autograph and needs no separate installation. Maxon supports this server for the connection between Autograph and your AI app.

If you set up a third-party server in the same AI app, disable that server while you use the built-in one. The model then sees only one set of Autograph tools.

No. Maxon does not charge for the Autograph MCP Server, which comes with Autograph, which is also free.

You provide the AI app and the model the app talks to. If the model runs in the cloud, you need an account with that service. Any related fees are between you and the AI service (e.g., Anthropic or OpenAI). A model that runs on your own computer through an app such as LM Studio needs no account with an AI service.

At this time, Maxon officially supports Claude Desktop and ChatGPT Desktop, and you can set up either one directly from Autograph’s preferences. Other AI apps that support MCP should also work, though Maxon has not tested them, nor does it officially support them.

Every AI app keeps a settings file that lists the servers the app may connect to. For the apps in the list, Autograph writes its entry into that file for you. For any other app, choose Other in the MCP Client list and click Copy Config. Autograph copies its entry to the clipboard so you can paste the entry into your app.

Yes. Apps that run models on your own computer, such as LM Studio, also work. Nearly any recent model can handle basic Autograph tasks, and stronger models usually give better results.

Note: A local model shares your graphics card with Autograph. A model that barely fits in graphics memory leaves nothing for rendering, and a slow model can take a minute to create a cube.

You need the full Autograph application (version 2027.0 or later) and the desktop client for your AI application of choice, both on the same computer. Nothing else is required.

The clarity of your request usually decides the quality of the result.

• The model. Basic tasks such as renaming work with almost any current model. Tasks that need reasoning about your project do not, and the gap between a small model and a strong one widens fast.
• The effort or reasoning level, if your app offers an effort setting. Low effort is fine for a quick rename. Give a bigger project setup more effort. That said, sometimes a medium setting beats a high one, because some models overthink. Test accordingly.
• The request. Describe the project elements, the frame range, the renderer, and how you plan to edit the result afterwards.

Note: A demo you see online rarely shows the prompts behind the demo or how many attempts the result took. A task that lands on the first try is the exception. Expect a conversation, with your second or third prompt showing much improved results. 

Most of that wait is the AI thinking, not Autograph working. Autograph is busy for only a fraction of a second at a time.

For added context, a task is a loop. The model reads your request and picks a command. Autograph runs that one command, usually in a split second, and sends back the result. The model reads the result, thinks, and picks the next command. The loop often repeats dozens of times, and the minutes go into the thinking. The status bar in Autograph shows which command is running. If a task drags, look at the AI app first.

While the MCP server working on a task, it is best to monitor the activity and progress of the AI, but without also using the computer and Autograph. In many cases, your actions can conflict with the MCP server tasks, or prevent the LLM client from completing processes. It is best to wait for the task to be completed, or pause the task, if you wish to take control.

A model gets answers from training data that may predate the feature, and some models even claim that no official Autograph MCP Server exists. Think of the situation as Schrödinger’s MCP. Support only exists when the AI observes it.

When an app connects, the server tells the model where the Autograph manual and SDK documentation live. A model confident in its training data may ignore that hint, so you may need to ask the AI directly to refer to up-to-date documentation. The model needs web access through the AI app to check the site.

The AI may describe its own work with more confidence than the work deserves. Check the X and the X before you accept that something is fixed or finished.

Only AI apps you set up can connect to Autograph. Unless you turn on remote connections, both sides of that MCP connection also have to run on the same computer.

Picture a lock and digital key. On first launch, Autograph makes a lock for itself and a matching key. Each AI app you set up gets a copy of that key, and anything without a key gets blocked. Autograph delivers the key for you, with no copying or pasting on your part. Keep the key private. Don’t write the key down, and don’t put your AI app’s settings or your Autograph preferences folder somewhere others can reach.

Once an AI app is connected, though, what it does with your scene is up to that app and its provider - check their settings if you want to know how they handle your data.

Your AI app asks, not Autograph. Each app decides how often to ask and whether you can switch the prompts off. Claude Desktop, for example, has an Auto mode at the bottom left of the prompt field.

Autograph controls something different. Under Edit > Preferences > MCP, you choose which tools the AI is offered at all and whether the AI may run Python.

In the Toolset section, under Edit > Preferences > MCP. Each checkbox controls one group of tools.

Details: The full list of tools and what each tool does is in a reference file in the MCP folder inside your Autograph preferences folder (Edit > Preferences > Open Preferences Folder…). Every command the AI sends is written to a log file in the same folder for auditing and safety, and the status bar shows each call as it runs.

The two Python preferences focus on security and logging. By selecting “Allow exec_python,” you let the AI assistant run its own Python code in the session for any functions the other tools don’t cover. The Verbose Logging feature records all Python that was executed and the full arguments and results of every call.

Yes. Each batch of changes can be reversed with Autograph’s undo functionality.

If a batch fails partway through, the steps that already ran stay in the scene and the rest are skipped.

Yes. You can disable the Autograph MCP Server at any time. Simply uncheck Enable MCP Server under Edit > Preferences > MCP, or run Stop MCP Server from the Extensions menu or the Commander. Connected apps lose their connection right away, and Start MCP Server brings the connection back.

Only if you allow remote connections. By default, the server only accepts connections from your own computer.

Opening the server to other computers takes three settings under Remote Access. Turn on Allow Remote Connections and confirm the warning, set Host to an address the other machine can reach, and add that machine’s IP address (or a range such as 192.168.1.0/24) to Whitelisted IPs. Don’t stop after the first setting. Turning on remote connections alone doesn’t let anyone in, because a request from an address not on the list gets refused. Browser-based apps also need their address listed under Browser Client Origins. Remote connections are not encrypted, so use remote access only on a network you trust.

Maxon Capsules are high-quality models, materials, and more that can greatly enhance your project. The Autograph MCP includes skills to search for and include assets from Maxon's Capsule library within your projects.

Keep in mind that while Maxon Capsules are licensed for end-users to include in commercial projects, they cannot be used to train an AI. To ensure your scene data is not used to train the AI model, you can often disable a setting in the AI tool's preferences. Before connecting an AI provider, you should carefully review that AI provider’s data usage and training policies.

For more details, please refer to Maxon’s End User License Agreement.

In general, the work an LLM does within an MCP on your behalf results in fully editable Autograph projects, but the LLM may not necessarily construct things in a way that makes them easily editable.

If you plan to refine a result yourself, saying so in your request can help make this easier for certain tasks. Generated animation, for example, may carry a keyframe on every second frame with linear interpolation, which is hard to shape by hand afterwards.

The answer depends on how the AI built the result. A rig driven by generated Python is hard for anyone to change without the conversation that produced the rig. Where handoff matters, ask the AI for a setup that uses standard Autograph features instead of Python.

Ask your AI app to create a simple shape, such as a star, in Autograph. If the shape appears, you’re connected.

Details: Most apps also list Autograph among their connected servers. In Claude Desktop, look for autograph under Settings > Connectors, with a check mark in the Status column. The server also appears under the + button in the chat box. Connection status and logs are under Settings > Developer.

This is usually because Autograph wasn’t running when the AI app started. Quit the AI app and restart it.

If restarting doesn’t fix the problem, check two more things. The Autograph MCP Server ships switched off by default, so make sure the Enable MCP Server setting is checked (or run Start MCP Server from the Extensions menu or the Commander). And if you clicked Regenerate Token, most apps stop working until you select each app in the Clients section and click Update Selected Client again, because those apps keep their own copy of the old key.

Most apps try to connect when they start and wait up to 20 seconds for Autograph. Once connected, an app reconnects on its own if you restart Autograph later, but the app won’t retry a first connection that failed. Some apps start at login and keep running in the background after you close their windows. Opening the window again doesn’t restart the app.

Autograph adds one entry for itself and leaves everything else in the file unchanged. Restart the app completely so the app reads the new settings.

Autograph saves a backup of the file before writing. If your app isn’t in the list, or Autograph can’t find a listed app’s settings file, choose Other, click Copy Config, and paste the entry in by hand.

No. Autograph touches only the entry Autograph wrote itself.

Every other entry is kept, the file is backed up before each write, and an entry Autograph didn’t write is reported and left alone. This applies to both Update Selected Client and Remove from Client.

The AI app can’t see Autograph’s dialogs and may carry on as if nothing happened. Dismiss the dialog, tell the AI what the message said, and ask the AI to check the scene before continuing.

Note: The status bar reports the failed call. A single failed call is normal, since a model often tries something, fails, and follows up with a corrected call.

No. Autograph MCP Server needs the Autograph interface and runs only in the full application, not in Commandline or Team Render clients and servers.