bitrouter-guardrails

GitHub repository: bitrouter/bitrouter

Local firewall for AI agent traffic at the proxy layer.

This crate provides a guardrail engine that inspects content flowing through BitRouter and enforces configurable rules. It wraps any LanguageModelRouter transparently via GuardedRouter, so the rest of the stack remains unaware of the filtering layer.

Includes

• Pattern-based content inspection engine in engine
• Built-in patterns for API keys, private keys, credentials, PII, and suspicious commands in pattern
• Configurable per-pattern actions (warn, redact, block) in rule
• User-defined custom patterns with regex and direction control in config
• GuardedRouter wrapper that applies guardrails to every routed model in router

Configuration

Guardrail rules are configured under the guardrails key in bitrouter.yaml:

• upgoing applies to outbound traffic (user/tool -> model)
• downgoing applies to inbound traffic (model -> user/tool)
• actions can be warn, redact, or block
• disabled_patterns turns off built-in detectors you do not want
• custom_patterns adds your own regex-based firewall rules

guardrails:
  enabled: true
  disabled_patterns:
    - pii_phone_numbers
  custom_patterns:
    - name: internal_ticket
      regex: "INC-[0-9]{6}"
      direction: both
  upgoing:
    api_keys: redact
    private_keys: block
  downgoing:
    suspicious_commands: block
  custom_downgoing:
    internal_ticket: warn

This lets you quickly define a local firewall policy for secrets, credentials, PII, or custom patterns without changing any application code.