README
Muzzle
muzzle/README.md
← Back to products
v1 v1
Product overview
PreviewMuzzle
AI input, output, and execution security layer
Our mission
Safeguards for every LLM, built for everyone
We're building Muzzle so anyone can decide what their language models are allowed to take in and give back — and genuinely trust the applications they build on top of them. Putting safeguards around an LLM shouldn't require a security team or custom infrastructure.
Muzzle makes it simple to set clear, custom rules around any model: filter the inputs, filter the outputs, and see exactly what was allowed or blocked.
Plain English
What Muzzle does
Muzzle is a transparent security proxy for AI model traffic. It sits between your applications and your model providers — Ollama, OpenAI, and Anthropic — and inspects every request, response, and tool call against your policy, with no changes to your client code.
- Runs in front of your provider on the port your clients already use.
- Inspects prompts, responses, and tool calls for prompt injection, secrets, and PII.
- Per-category actions — allow, log, redact, transform, or block — with tool calls blocked by default.
- Fails closed by default: if inspection cannot complete, traffic is blocked rather than passed through.
Install
Installs from source on a Linux host (needs python3.12 and sudo).
curl -fsSL https://ai.hurricaneindustries.co/api/products/muzzle/artifact -o muzzle.tgz
mkdir -p muzzle && tar -xzf muzzle.tgz -C muzzle
cd muzzle && sudo bash install.shOperational summary
Visibility
public
Downloads
Available
Latest release
v1
Working version
v1
Living documents
Browse historyv1
PLAINENGLISH
Muzzle in Plain English
muzzle/PLAINENGLISH.md
HOWITWORKS
How Muzzle Works
muzzle/HOWITWORKS.md