Local connections in the desktop app

The desktop app can start local MCP servers via stdio. This gives the AI access to extended Google and Microsoft features, Nextcloud, and other tools that a browser cannot run. Local file access is built-in and does not require an MCP server.

Onboarding

During the initial setup, you go through these steps:

1. You choose a local default folder for the built-in file tools.
2. You can optionally choose a local connection for Nextcloud.
3. You can optionally add other MCP servers from the tool store.

You can postpone the entire onboarding. The default folder and any additional connections can be set later. All folder choices are stored locally only.

The tool store

Click Add tools or Add MCP server. You will first see safe presets that fit the desktop app. Only when you choose Add your own MCP server does the manual form open.

Common presets include:

Preset	Capabilities	Authentication
Nextcloud	Files, calendar, and contacts	Server address, username, and app password
Memory	Save and retrieve knowledge later	Local storage

The store may also show presets for services such as GitHub, Canva, Notion, Supabase, Mailchimp, and other suitable MCP servers. Availability depends on platform, authentication, and the current catalog.

Google and Microsoft

Google Workspace and Microsoft 365 are built-in AI-School connections. They work on web, mobile, and desktop and do not depend on OpenAI or a local MCP preset. After one-time Google or Microsoft consent, the available tools can be used with OpenAI, Claude, Gemini, and Mistral, as far as the chosen model supports tool usage.

Dynamic tool selection

An MCP server can offer dozens of tools. Therefore, the Desktop Deep Agent first loads compact tool information and selects the relevant tools during the task. This way, multiple large MCP servers can be active without placing all tools in the model context at once.

Secure installation

Local MCP installation has a strict boundary between the Angular interface and Electron:

• the renderer does not start arbitrary shell commands;
• command, arguments, environment keys, and requested access are checked;
• a native dialog shows what is being started;
• approved configurations receive a proof of integrity;
• secrets are encrypted with the operating system’s secure storage;
• processes stop on disable, removal, logout, and shutdown.

If you remove a server, the locally stored credentials are also deleted.

Status and testing

Under On this computer you see all active local servers in one place. In the chat settings, local and remote connections are grouped together under Connections.

Test a server with a concrete question, for example:

• “Which folders are in my Nextcloud?”
• “Find the latest emails about the project.”
• “Which files are in this local folder?”

If there are problems, check the desktop log. A successful connection mentions the server name and the number of loaded tools.

Troubleshooting

• Server is active but the AI sees no tools: restart the local runtime and check if the server ID is in the chat request.
• Authentication missing: open the server in Preferences → Tools and choose reconnect.
• Nextcloud does not work: use the base URL of the Nextcloud environment and a separate app password.
• Local files are not accessible: check if the active working folder still exists and reselect it via the dashboard if necessary.
• Too many tools: keep dynamic tool selection enabled and clearly specify in your question which source is needed.

Adding your own stdio server

Preferably use a preset. Only add a server manually if you trust the origin and the command. Before approval, check:

• executable and arguments;
• environment keys;
• filesystem access;
• read and write capabilities.

See also

• Connections and MCP
• Add your own MCP server

Access to local folders

The built-in file tools can only read or modify files within the working folder that you explicitly approve via the native folder picker.

Each path is resolved to the canonical location. Tricks with a similar prefix of a neighboring folder and symbolic links that lead outside an approved folder are rejected. After restarting the local runtime, permissions are resynchronized; without an approved folder, file access is denied by default.

If you ask the AI to delete a file or subfolder, the desktop app moves the item to the system recycle bin. This works on Windows, macOS, and Linux. The approved main folder itself cannot be moved to the recycle bin via the AI.