MCP Drift Watch

For teams rolling out MCP

Know when an approved MCP server changes what it tells your AI.

AI agents follow tool descriptions like instructions. Drift Watch fingerprints every server you approved, and freezes it the moment its tools change, until a person re-approves.

uv tool install mcp-drift-watch

or pipx install mcp-drift-watch · free · Apache‑2.0

Terminal and review site: a changed MCP server is caught, frozen and kept away from the model.
A test server pulls the rug: Drift Watch catches the change, freezes the server, and the gateway keeps the new tools from the model.

The rug pull nobody sees

You checked a server before approving it. Your AI uses it every day after that. Its tools can change anywhere in between.

Approved once, trusted forever

An MCP server can change its tool descriptions any time after you approve it, with the same name and the same version number. Nothing in MCP tells you.

Instructions in plain sight

Your AI reads tool descriptions as instructions. A changed one can tell it to read files or send data elsewhere, in text a person may not even see: invisible characters, a new parameter.

Hit and run

A server can serve a poisoned list for a few minutes, then switch back. Apps keep the poisoned copy cached for up to 24 hours, while the server looks clean to anyone who checks.

How it works

Drift Watch compares what each server says about its tools with the version a person approved. It never needs your servers' source code.

1

Watch

Every approved server is pinned by a SHA-256 fingerprint of what the AI reads. The watcher re-checks by risk, from every 6 hours to weekly, or right now. Anything changed is frozen, and Slack and email get an alert with its OWASP MCP Top 10 risks.

2

Decide

A person sees the approved and new versions side by side, with hidden text in red. Claude gives a second opinion, quoting the risky text. Only a person can approve.

3

Enforce

The guard, in agents you write, and the gateway, for apps you can't change, pass on approved tools only. That includes copies in an app's cache, where a hit-and-run hides.

Measured, not promised

From our own test servers and test changes.

0 of 20harmless changes raised a false alarm with the AI reviewer. Hash diffing alone: 12 of 20.
20 of 20meaningful changes caught, on 40 test changes Drift Watch was never tuned on.
31.5 sfrom a pushed rug pull to the Slack alert, with “re-check now”.
≈1 sfrom an app’s first contact with a poisoned server to the alert, through the gateway.

The test changes are synthetic, so these are upper bounds, not guarantees. One limit we're open about: Drift Watch sees what a tool says, not what it does. A change in a server's code alone isn't caught.

What it never does

A security tool has to be safe to run.

Pricing

The CLI is free, for good. Team and Enterprise are in early access.

Free

The CLI, on your machine

$0

  • Pin and check any number of servers
  • Diffs with red flags and OWASP IDs
  • AI second opinion with your own key
  • Guard and gateway
  • Review site, on your machine
Install the CLI

Enterprise

For the whole company

$499 / month

  • Unlimited watched servers
  • Single sign-on (SSO)
  • Audit export
  • Everything in Team
Join the waitlist

Get early access

Team and Enterprise open to the first teams on the list.

I'm interested in

We store your email address, the plan you picked and the time, and nothing else. No cookies, no tracking, no analytics. We'll use it only to tell you when your plan opens.