How to choose where your AI works on a project

Published on November 04, 2026

When you connect Codex, ChatGPT, or Claude via MCP, you must talk to that agent in its own app: Sawdust's in-app chat only talks to Dusty and cannot reach your MCP agent. To give an AI agent control inside Sawdust's chat panel, add a model with an API key.

When you connect an external assistant such as Codex, ChatGPT, Claude, or Cursor to Sawdust via MCP (https://sawdust.diy/mcp), you have to work with that agent in its own app or terminal. The MCP connection lets that external agent inspect and edit your Sawdust projects, but it is a one-way connection into Sawdust: the in-app chat panel inside Sawdust cannot send messages back to your connected MCP agent.

Inside Sawdust's chat panel, you can always chat with Dusty (Sawdust's built-in shop helper for ready-made plans and part edits), even while an MCP assistant is connected to your account. However, you cannot chat with your connected MCP agent from inside Sawdust, and Dusty cannot pass instructions to it. If you want to give an AI agent direct control inside Sawdust's in-app chat panel, you must configure a model using an API key.

1. Working with Codex, ChatGPT, or Claude via MCP

When you authorize an assistant through https://sawdust.diy/mcp, you use that assistant's own subscription and interface to drive your Sawdust builds. All prompting happens in the external app or CLI where you connected Sawdust:

1. Connect your assistant once using https://sawdust.diy/mcp and approve the OAuth consent screen in your browser. Follow the step-by-step guide for your client in Connect an AI assistant to Sawdust.
2. Open Codex, ChatGPT, Claude, or your MCP client and type your request there. Name the Sawdust project you want it to inspect or edit, for example: "In my Sawdust 10x10 shed project, add a ramp to the front doorway."
3. After the assistant finishes applying changes, reload the project page in Sawdust to see the updated 3D model, build steps, and cut sheet. An already-open Sawdust browser tab does not automatically refresh when an external MCP agent writes to the project.

To check which MCP assistants currently have access to your account or to revoke a connection, open Settings -> Connected assistants in Sawdust.

2. Giving an AI agent control in Sawdust's in-app chat (API key)

Consumer chat subscriptions (such as ChatGPT Plus or Claude Pro) only allow you to chat inside their own apps and connect outward via MCP; they do not let third-party apps run the model inside another chat interface. If you want an AI agent to work directly inside Sawdust's chat panel, where it can see your live 3D viewport, take part selections from the scene, and update your model in real time as you watch, you must add a model with an API key:

1. Click Add your AI model beneath a Dusty reply in the chat panel, or open Settings and click Add your own AI agent (labeled Manage your AI agents once you have saved a configuration).
2. In the Your AI agent dialog, select Use an API key.
3. Choose Anthropic or OpenAI-compatible under Dialect, confirm the Endpoint and Model, paste your provider API key, and click DONE (or Add config).
4. Send your build or edit requests directly in Sawdust's chat panel. When both Dusty and your API-key model are enabled, Dusty handles standard ready-plan and part operations while your model takes over custom designs, procedures, and build videos. To send every message straight to your model, select its name on its own in the chat header's model selector.

API usage is billed directly by your model provider (such as Anthropic, OpenAI, or OpenRouter) based on token usage, separately from any monthly chat subscription. For more on what Dusty handles on its own and when to bring in a custom model, see Using Dusty and When to connect an AI model.