Muzzle stops Claude Code and Codex from sending out your email, keys and secrets, and catches pages and files that try to take them over.
MuzzleStopped going out
Stopped coming in
Commands and files are checked before they run or get written.
A commit that would leak doesn't go through.
A local proxy checks the requests that scripts your agents run actually make.
At setup, a scan of your code shows what already leaks.
Pushing to GitHub, deploying to Vercel or publishing to npm goes through as usual.
A Stripe key going to Stripe passes. The same key going to webhook.site is blocked.
sk_live_…api.stripe.comAllowedsk_live_…webhook.siteBlockedFor the first week Muzzle only watches, then shows you what it would have blocked so you can turn blocking on.
Pro
Crawlers, APIs and push keys often want a contact email. Muzzle gives you a personal address that forwards to you, and when it blocks your real email it tells the AI to use this one.
Macs that contribute to the Threat Feed report what they catch. It becomes a signed rules list that every Muzzle install downloads daily.
Injection seen on many Macs
Flagged on all
Cleared on many Macs
Warns instead of blocks
Trusted on many Macs
Suggested to you
$8/month, billed $96 yearly
$19/month
+ VAT where applicable
Free when you contribute to the Threat Feed
Claude Code and git on Free. Pro adds Codex, and checks what the scripts your agents run send through the local proxy.
The email, phone number and name you enter at setup, plus API keys and tokens from Stripe, AWS, GitHub, OpenAI, Anthropic, Slack, Google, Supabase and others, and private keys.
The checks run on your Mac, and traffic to your trusted platforms passes straight through.
Update checks, the daily rules download and, on Pro, licence checks. Free needs no account; Pro through the Threat Feed needs a confirmed email. The contact address needs your real email to forward mail to, and the Threat Feed collects only what's listed below. Privacy has the details.