Remote servers
For a remote MCP server, Garm still runs as a stdio server for your agent and talks streamable HTTP to the server. The same policy, pins, audit log and shared session apply, and a call Garm stops never leaves your machine.
{
"mcpServers": {
"github": {
"command": "garm",
"args": ["mcp-proxy", "--policy", "garm-policy.yaml", "--name", "github",
"--url", "https://api.githubcopilot.com/mcp/"],
"env": { "GARM_HEADER_1": "Authorization: Bearer ${GITHUB_TOKEN}" }
}
}
}
garm init writes this for servers that are configured by URL.
Headers and tokens
Put headers in GARM_HEADER_* environment variables, as Name: value. Garm expands ${VAR}
from its environment and never logs header values.
Avoid --header arguments for tokens: agents expand ${VAR} in a server’s arguments before
starting it, which puts the token on Garm’s command line, where other users of the machine can
see it.
Details
httpsonly, except plainhttptolocalhost,127.0.0.1or[::1].- TLS uses your operating system’s certificate store, so corporate CAs work.
- Each message is its own request, so a slow tool call doesn’t hold up the next.
- If the server can’t be reached or doesn’t answer, the agent gets an error for that request instead of waiting.
- Not supported yet: OAuth sign-in for remote servers (use a token header), resuming dropped event streams, and the older SSE transport.