Vigils

Vigils is a local control plane for AI agents, built with Rust, Tauri and SQLite and licensed under Apache 2.0. Its stated purpose is to show what agents do, require approval for risky actions and keep credentials out of model context. It sits in front of an agent's tool calls and credential flow, applying policy before anything runs. It is aimed at developers and teams who use agents such as Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Kimi CLI, ZCode and pi and want oversight and an audit record. Claude Code, Codex, Gemini CLI and Cursor connect through hooks, and the others through an MCP gateway. The core mechanisms include a policy engine with a deny, then approve, then allow ordering, a tamper evident SHA 256 audit chain with full text search, and credential leases that inject short lived secrets so real credentials never enter the model context. An approval queue holds high risk operations for a human decision, with once, this session and for tool scopes. Agent requests can run in a Wasmtime sandbox with the environment cleared and network access denied, and a PII redaction layer combines hard rules with an ONNX model to scan for eight types of sensitive data. It comes in three forms: a desktop application that stays in the system tray, a Chrome extension that intercepts pastes and submissions on ChatGPT, Claude, Gemini and Perplexity, and a command line toolset for serving the MCP gateway, searching audits and replaying sessions. The site lists no pricing.