Muzzle in Plain English v2
This is the stored snapshot for the approved document version. The diff below shows what changed from the previous version.
Document snapshot
Muzzle in Plain English
Summary
Muzzle helps stop AI systems from accepting dangerous instructions or leaking things they should not reveal. It sits quietly between your apps and the AI service and checks everything going in and coming back out.
How you'd use it
You point Muzzle at the same address your app already uses to reach the AI. Nothing in your app changes. Behind the scenes, Muzzle reads each request before it reaches the AI and each answer before it comes back, and applies your rules.
For each kind of risk — someone trying to trick the AI, secrets, personal information, or banned words — you choose what should happen: let it through, just make a note, hide the sensitive part, rewrite it, or block it entirely. If Muzzle ever can't finish its checks, it blocks by default rather than letting something slip through. Requests that ask the AI to run a tool or take an action are blocked by default.
Current build
Muzzle v1 is in preview. It works in front of Ollama, OpenAI, and Anthropic, treating all three the same way so your rules don't change when you switch providers. It also understands streamed (typed-out) answers and tool/action calls, not just plain replies.
Long answers used to be a problem: if an AI took several minutes to think, something in the chain would give up on it and the whole answer was lost even though the model was still working. Muzzle can now read the answer as it is being written and put it back together, and how long to wait is set per AI service rather than once for everything — a small local model and a big remote one get their own patience.
An installed copy comes with:
- A small admin screen for reading and editing the settings as a simple form, trying out rules before you commit to them, and watching a live log of every decision. You can add or remove rows — providers, ports, per-provider rule sets, and banned words — right in the form.
- A
muzzlecommand for routine changes: adding a provider, adjusting a rule, checking the settings are valid, seeing status, previewing a decision, and restarting the service. - An uninstaller that cleanly removes everything it set up.
Where the words on the website come from
The product write-ups in this repository are the source of truth. When a set of docs is approved, a copy of each one is saved into the website's database with a timestamp, so the site always shows the latest approved version and can also show older versions and what changed between them.
Diff from previous
--- PLAINENGLISH@1 +++ PLAINENGLISH@2 @@ -26,6 +26,12 @@ understands streamed (typed-out) answers and tool/action calls, not just plain replies. +Long answers used to be a problem: if an AI took several minutes to think, something +in the chain would give up on it and the whole answer was lost even though the model +was still working. Muzzle can now read the answer as it is being written and put it +back together, and how long to wait is set per AI service rather than once for +everything — a small local model and a big remote one get their own patience. + An installed copy comes with: - A small **admin screen** for reading and editing the settings as a simple form,