Quick start
Run these in a project that has an MCP config: Claude Code’s .mcp.json, Cursor’s
.cursor/mcp.json or VS Code’s .vscode/mcp.json. For another file, such as Claude Desktop’s
claude_desktop_config.json, add --config <file>.
1. Assess
garm assess
For each server, Garm shows what it reads and where it sends data, then whether one session can hold all three parts of the lethal trifecta: untrusted content (the vector), sensitive data (the payload) and an external sink (the egress).
Lethal trifecta: EXPOSED
Vector (untrusted content): github
Payload (sensitive data): postgres; fs if it returns secrets
Egress (data leaves): github, fetch
For example: an injection in github content, then a read from postgres, then a call to github.
2. Put every server behind Garm
garm init --dry-run # preview
garm init
garm init rewrites each server to run behind garm mcp-proxy, matches the servers it
recognises to label packs, and writes garm-policy.yaml in observe
mode: Garm logs what it would hold or block but lets everything through. It backs up your
config first; garm init --undo restores it.
3. Run your agent under Garm
garm run -- claude
garm run hosts one session for the whole agent run. Every proxy the agent starts joins it, so
content read through one server constrains calls to another. Without it, each server gets its own
session and cross-server leaks aren’t caught.
4. See what Garm would have stopped
garm audit tail
When the log looks right, set mode: enforce in garm-policy.yaml.
Before you rely on it
Garm only sees traffic that goes through it. Your agent’s built-in tools (reading files, running
shell commands, fetching URLs) don’t go through MCP, so Garm doesn’t see them. Turn them off
where you can (for example Claude Code’s --disallowedTools), so file and web access goes
through MCP servers behind Garm. See limitations.