← Back to Muzzle
Internal lab sheet
PreviewMuzzle v1 technical sheet
internal lab sheet for the installed v1 build. It describes the runtime, listeners, policy flow, and operator workflow in technical terms.
Runtime
Single Python process with two listeners
Proxy listener
0.0.0.0:11434
Admin listener
0.0.0.0:11435
Failure mode
Fail closed by default
System detail
Architecture
- Transparent proxy in front of Ollama with a rules-first inspection path.
- Same process hosts the proxy listener and the local admin listener.
- Configuration is YAML-driven and loaded at startup from the installed config file.
Request path
- Inbound prompts are inspected before they reach the upstream model service.
- Streaming responses are buffered so the output policy can run before release.
- The proxy preserves the client-facing Ollama shape unless a policy decision blocks the request.
Policy pipeline
- Prompt-injection, secret, PII, and content-policy checks run first.
- Category actions can allow, log, redact, transform, or block.
- Inspection failure is treated as a rejection unless the config explicitly changes fail mode.
Admin UI
- The local admin listener exposes the live config and the decision log.
- Config edits are written back atomically and reloaded into the running process.
- Log view supports tail, time, request, stage, upstream, action, category, and text filters.
Install and runtime
- The Linux installer writes the default config, systemd unit, and log directory.
- Proxy traffic listens on 11434; the admin UI listens on 11435.
- The default install is preview-only and does not publish a download artifact.
Current scope
- Targets Ollama today.
- Preview release only.
- No public download link is exposed from the product page.
Internal use only. This page is for lab operators and deployment work. There is no public download from Muzzle v1.