A local model can write an answer, but it needs a connection to read current information from your apps. OGAD (Off Grid AI Desktop) supports MCP, a standard that lets a server expose app tools to the assistant. You can connect a supported service or a local server, then ask the model to use its available tools.
Basic custom MCP connections are in the core app. Some built-in integrations need Pro or additional provider setup. An integration marked Not enabled yet is not a working connection to try.
What do you need?
Use a local text model that works with tool calls, plus a working MCP server for the app you want. Get its connection details from that server's documentation. A service login page or normal API address is not necessarily an MCP endpoint.
A local server must already be installed with its required runtime. A remote server needs its network connection and authentication. MCP does not make a cloud service offline.
Add a custom connection
- Open Integrations and select Custom.
- Enter a Name that identifies the service.
- For a remote endpoint, choose HTTP / SSE and enter the documented MCP URL. For an installed local server, choose stdio (local) and enter its command and arguments.
- Select Add, then open the connection and select Test. Complete authorization if requested.
- Check that the connection is on. In chat options, enable Tools and Connectors.
- Ask for one small read operation that the server supports, such as listing a few documents in a test folder.
Use the server's documented command. The local command field starts a program on your computer; do not paste an unfamiliar install command merely because a page calls it an MCP setup.
Check the first answer against the app
Name the service and the exact object you want in the request. For example: “Use my connected notes service to find the note titled Test checklist and summarize its first three items. Do not change it.”
Open the source note and compare the answer. This checks both the selected connection and the model's use of the returned data. Do not assume that a successful connection test proves every tool works with your account.
If a tool is missing, check what the server actually exposes. If authorization fails, reconnect with the required account and permissions. If the model answers without calling the service, check the chat's Tools and Connectors controls.
Where does the data go?
A local model processes the tool results on your computer. The connected server still receives tool requests and can access data according to its permissions. A local server may itself contact a cloud service. Check that server's behavior before calling the workflow fully offline.
Some connector tools can change data directly from a chat request. Start with a read-only task and be explicit about whether you want a draft or an actual change.
Try OGAD with one trusted MCP server and one test record. Confirm the source result before you use the connection for a larger workflow.
