Virgil

Virgil is a local circuit breaker for supervised AI agents. One executable runs the core, gateway, and browser panel on the same machine. A policy block is recorded and terminates the supervised process tree. The panel is available at /dashboard on the loopback address. The gateway supports configured OpenAI-compatible requests, provider adapters, local policy checks, SQLite event and execution history, and JSONL export. The core binds to the configured local address. Setup is served locally when the configuration is missing. The local product does not require an account or hosted Control Plane. It currently binds to loopback; secure LAN operation is a separate future stage. The gateway can block only traffic routed through it and does not terminate an external agent process. The current proxy supports JSON text messages, tool calls, and SSE streaming. Unsupported fields return an error before provider dispatch. Provider type may also be openai, kimi, or anthropic. OpenAI, Kimi, and generic endpoints use the configured OpenAI-compatible protocol; Anthropic translates the supported text/tool subset to Messages. Guardrails are evaluated locally before a request is sent, while a stream is active, and after tool calls are observed. Initial policy types include maximum cost per execution, maximum input tokens, maximum output tokens, maximum total tokens, maximum number of provider calls, maximum execution duration, maximum tool calls, allowed providers, allowed models, and allowed tools.