Repository navigation
Numbat
Teams adopting AI coding agents have no independent record of what those agents do on a machine. numbat is a Go single-binary command-line tool from Perplexity that observes desktop, CLI, IDE and gateway agents, including Codex and Claude Code, through hooks, plugins and OpenTelemetry (OTLP) logs. It runs detection locally with a CEL rule engine over normalized events and offers forensic reconstruction of past agent activity.
The tool addresses a gap in agent adoption: an agent can read files, run commands and change systems, yet the organization running it has no audit trail and no way to intervene. numbat gives that trail and an optional control point. Detection rules are shipped as YAML files. Pre-action blocking is off by default and is enabled per rule by setting enforce: true, so adoption can start as pure visibility and tighten later.
Its limits are stated by the project. At-rest reconstruction cannot recover activity that was never persisted, and a finding is a rule match, not proof of compromise. The repository was created on 2026-07-24, has about 1.1k stars and 19 open issues, and supports macOS, Linux and Windows on amd64 and arm64.
It overlaps in purpose with Agenttrace and AgentGovernanceToolkit, which address agent observability and policy from other angles.
numbat sits in Assess. It is backed by a well-known vendor and is under active development, but it has no first-person use behind it and its detection quality is unverified in practice.
