Local MCP servers · DeutschlandGPT

A local MCP server is a program that runs on your computer and offers tools to the assistant, for example access to a folder, a Git repository or a local database. Code and Cowork can then call those tools.

You manage local servers under Settings, Integrations. In Code and Cowork, /mcp opens the integrations directly, filtered to MCP servers.

A local server runs with your user rights on this device, and what its tools return is sent to the selected AI model. Only add servers you trust. The app asks you to confirm this before it adds one.

Local or remote

The list under Integrations has two areas:

Connected to your account: the prebuilt integrations, your own remote MCP servers and WebDAV. A remote MCP server runs on the internet, and our server connects to it. That is why it is available in every product, in the browser as well as in Chat, Code and Cowork.
On this device: your local MCP servers with the Local badge, their state and the hint Only in Code and Cowork. Below them, under Config files, are the configuration files the app found and their servers, including those that are Not approved yet.

The filters All, MCP servers and Local narrow the list, and the search finds entries in both areas.

Adding a server

Under Integrations, click Add. The app first asks what you want to add: a Remote MCP server (an address), a Local MCP server (a command) or WebDAV / Nextcloud. Above these three cards is Paste configuration. If you have a configuration with several servers, use it: the app accepts JSON with mcpServers (or servers as in VS Code), a single server entry or a command line, and sorts each entry on its own. Entries with a command become local servers, entries with a URL become remote MCP servers.

For a local server there are four ways, and none of them starts anything until you have saved:

Popular servers: one click adds a ready-made entry, for example File system, Git, SQLite, PostgreSQL, Web fetch, Memory, Browser (Playwright) or Obsidian. You only choose the folder, file or connection URL it needs. If the server needs Node.js or uv (Python) and the app cannot find it, it shows the install command.
Found in other apps: servers already set up in Claude Desktop, Claude Code, Cursor, VS Code or Windsurf are offered with Take over. Entries with a URL in those files are offered as remote MCP servers.
Paste: paste the JSON from another app (with mcpServers) or a command from a README, such as npx -y @modelcontextprotocol/server-filesystem ~/Documents. An entry with a URL opens the form for a remote MCP server, and the pasted list is still there when you come back.
Add manually: enter a Name, the Command, its Arguments (one per line), Environment variables, an optional Working directory and a Time limit per call (seconds). Test checks whether the server answers and how many tools it offers.

The menu More options (⋯) next to Paste and Add manually has two more entries: Open config file shows the file in which the app keeps your local servers, and Export copies them as JSON with mcpServers, without the values of the variables.

The name becomes the prefix of the server's tools: lowercase letters, digits and hyphens, at most 24 characters. The command is started directly, without a shell.

Managing servers

Each local server shows its state in the list: Running, Stopped, Starting… or Error. Clicking it opens its details. There you switch the server on or off, start or stop it, and use the switch next to each tool to offer only the tools you need. read-only marks tools that declare they change nothing. Log shows what the server printed, Edit changes its settings and Remove deletes it.

At most 40 tools are offered to the assistant at a time, across all servers. The app warns when you have switched on more.

Servers from a configuration file

As in Claude Code, you can also list servers in a file. .deutschlandgpt/mcp.json or .mcp.json in a project folder applies to that project, ~/.deutschlandgpt/mcp.json to all projects. A server from such a file only starts once you have reviewed and approved its command. If the entry changes, the app asks again.

Where the settings are kept

Local servers and their settings stay on this device and are never synced to our server. Values of environment variables are stored in your operating system's keychain and are not shown again after saving. Removing a server also deletes its stored values.

Approvals

A call to a local tool goes through the same approvals as everything else in Code and Cowork:

In Plan mode only tools marked read-only may run.
In Auto mode tools run without asking. A sub-agent still asks.
Otherwise the app asks before a tool runs. If you allow it permanently, that permission applies to this tool with exactly this server configuration. If you change the command or its arguments, the app asks again.

Calls of tools that may change something are recorded in your organization's audit log with server name, tool name and result. Arguments, values and outputs are never recorded.

For administrators

The switch Allow local MCP servers under Tool policies decides whether members may add local MCP servers in Code and Cowork. Every call runs with the member's rights on their device, and the results go to the selected AI model. When you switch it off, the desktop app stops all local servers at its next sync.

When local servers are not available

Message	Meaning
Not available yet.	Local MCP servers are not switched on for your account
Turned off by your organization.	An administrator has turned off Allow local MCP servers for the organization. Running servers are stopped