Stop AI agents from leaking your data.

Garm sits between an agent and its tools. It labels everything that enters the agent's context by where it came from, tracks those labels across the whole session, and stops the call that would send sensitive data out.

Early development · building in public · quick start

The lethal trifecta

An agent leaks data when one context holds three things at once: untrusted content (a web page, an issue, an email), sensitive data (secrets, customer records, private code) and a way out (an HTTP request, a comment, an image URL, a DNS lookup). A prompt injection in the first tells the agent to move the second through the third.

You can't patch prompt injection out of a model. Garm assumes the agent will be tricked and takes away the way out whenever the other two are present.

Session labels, not per-call scans

Per-call DLP sees a harmless-looking fetch. Garm knows the session read a GitHub issue and a .env file first, even through different servers.

Sees through encodings

Secrets and PII are found in base64, hex, URL-encoded and gzipped arguments, nested up to four layers deep.

Rug-pull protection

Tool definitions are pinned. If a server changes a tool after you approved it, Garm hides and blocks it until you review.

Signed audit trail

Every decision goes into a hash-chained, ed25519-signed log you can verify, without ever storing the secret itself.

Works with your agent

Claude Code, Cursor, VS Code, Claude Desktop: anything that speaks MCP, over stdio or streamable HTTP. No changes to the agent.

Fast enough to forget

Well under a millisecond added per tool call at p99 for local servers, measured nightly against a 5 ms budget. Written in Rust.

Try it on your agent

garm assess          # does your MCP config complete the trifecta?
garm init            # put every server behind Garm, in observe mode
garm run -- claude   # one Garm session across all servers
garm audit tail      # what Garm would have stopped

Read the quick start · all docs