$47 burned overnight because two AI agents couldn't agree on JSON formatting. That's the kind of problem Agent Firewall is designed to solve.
The tool is a Go-based reverse proxy aimed at anyone running multi-agent AI setups - think CrewAI, AutoGen, or LangGraph, where multiple AI models work together on tasks. The failure mode it targets is real and common: agents enter infinite loops, endlessly retrying or arguing with each other, while your OpenAI or Anthropic bill climbs by the minute with nobody watching.
Setup is minimal. You point your existing API calls through the proxy by changing a single environment variable (OPENAI_BASE_URL=http://localhost:8080/v1), and Agent Firewall monitors traffic patterns for signs of a death spiral - repeated similar requests, escalating token counts, circular conversation patterns. When it detects one, it physically cuts the connection.
The design choices are sensible: it runs locally with no cloud dependency, requires no SDK changes to your existing code, and works as a transparent proxy so your agents don't need to know it exists. That last point matters because most agent frameworks don't have built-in circuit breakers (a mechanism that automatically stops a process when it detects a failure pattern), and bolting one on after the fact usually means rewriting your orchestration layer.
This fills a real gap. Anyone who has left an agent swarm running overnight knows the sinking feeling of checking your API dashboard the next morning. Rate limits help, but they don't distinguish between productive work and an agent talking to itself in circles. A proxy-level kill switch that understands conversation patterns is a more targeted solution.
The project is early-stage and open source. If you're spending real money on multi-agent workflows, it's worth testing in a staging environment before trusting it with production traffic.