Evaluate Firefish
Firefish protects the boundary where prompts, retrieved content, outputs, streaming responses, and proposed agent actions become trusted application behavior.
Start locally
Run the gateway, scan one prompt, validate one tool call, and open the operator dashboard.
Review the proof
Inspect benchmark methodology, source-aware coverage, and the public proof center.
Doc map
- Quickstart - first scans, tool validation, and protected chat.
- Architecture - scanner, policy, runtime, tools, audit, and dashboard.
- Threat Model - in-scope risks and explicit non-goals.
- Benchmarking - recall, precision, FPR, latency, and source-aware evaluation.
- Tool Governance - action classes, approvals, destination risk, and goal alignment.
- Deployment Modes - local dev, Docker, private deployments, env vars, and production checks.
- Python SDK - client calls, FastAPI guard, and tool-call validation.
Public and protected surfaces
Marketing and documentation pages are public. The operator dashboard remains under /app/ and keeps the existing dashboard authentication behavior.