Four steps from install to a verified connection.
MCP Stub is a connection, not an application: you check it is installed, add one entry to your agent's configuration, and confirm it with a read. It needs no network setup of its own and no root access.
- # Expected command; run it to obtain your installed version.$ mcp-stub versionStep 1 of 4
Check that it arrived
The MAIT image ships
mcp-stubfor Linux on amd64 and arm64 and updates it with the image. On macOS and Windows, MCP Stub runs in the managed Linux environment that AIDC Launch provides. Check what you have.What you getThe program answers with its version. A build can carry a commit suffix, so report what you see.
Choice Use it when Setting An existing service The service owner already runs it on a Unix socket or an HTTP address --server ENDPOINTA registry on demand You own the host and want the stub to start a registry for you --registry-key KEYand--registry-json PATHStep 2 of 4Pick how it reaches the service
Choose one of two ways. They are mutually exclusive.
On a shared host, use a Unix socket in a private directory only your account can enter. Use only an owner-approved endpoint. Keep HTTP on loopback, or put an authenticated, encrypted proxy that you operate in front of it.
What you getOne decision, written down before you edit any configuration.
{ "mcpServers": { "mait": { "command": "/usr/local/bin/mcp-stub", "args": ["serve", "--server", "/run/user/1000/mcp-registry.sock"] } } }Step 3 of 4Add the entry to your agent
Any client that accepts an
mcpServerslaunch entry uses this shape. Write the socket path out in full: configuration files do not expand$XDG_RUNTIME_DIR.To have the stub start a registry instead, replace the arguments with
["serve", "--registry-key", "agent-1", "--registry-json", "/etc/mait/mcp-backends.json"]. Use absolute paths. Use on-demand startup only on a host where you are the only operating-system user. On a shared host, use an existing service with an owner-approved private socket.What you getA client that starts the stub on its own when it opens the workspace.
- Step 4 of 4
Verify one useful read
Open your agent in the workspace and enable the connection through its normal controls. Ask it to refresh its tools, then request a small read you are allowed to make, for example "Read issue PROJ-1234 and summarize its status", with your own identifier.
The expected result is the current content and a summary based on it. If the agent cannot name a tool for the task, or returns an access error, the connection is not verified. A visible tool name proves neither permission nor that the service is healthy.
What you getA connection proven by a real result, not by a list of names.
A connection that reconnects after a restart.
Your agent reaches its tools through the stub. If the service restarts, the tools stay listed. After it recovers, ask the agent to refresh its tools and repeat the small read before you rely on it. Before you repeat an interrupted change, check the destination: the first attempt may have completed even though its answer was lost.