Numbat

Numbat: Security software providing endpoint visibility into AI agent activity with on-device detection, optional pre-action blocking, and forensic reconstruction.

Numbat observes activity from desktop, CLI, IDE, and gateway AI agents using local hooks, OTLP/HTTP logs, and session artifacts. It normalizes this data into a single event model evaluated by a Common Expression Language (CEL) rule engine. The software runs detections locally and supports optional pre-action blocking for specific agents. It provides tools for forensic reconstruction from on-disk artifacts and inventory management. This tool is designed for security operators to monitor and control agent behavior without relying on cloud-based processing.

Numbat at a glance

Best for: Software engineer
Runs on: macOS, Linux, Windows, CLI
Pricing: Free
Released: July 2026
Category: AI coding tool
Website: github.com/perplexityai/numbat

Key features

Live monitoring via hooks, plugins, and OTLP/HTTP log exporters
Endpoint-local detection using built-in, multi-step, and custom YAML rules
Opt-in enforcement blocking through pre-action hooks
Forensic reconstruction from on-disk session artifacts
Versioned NDJSON record generation for events and findings
Read-only artifact scanning with secret redaction
Single-binary distribution without cgo

Use it for

Monitoring live AI agent activity for unauthorized behaviors
Forensically reconstructing past AI agent sessions from artifacts
Enforcing security policies on agent tool calls via pre-action hooks
Auditing AI agent configuration and discovery across endpoints

Alternatives to Numbat

Daytona — It provisions sandboxes in under 90 milliseconds with full security isolation.
TaintWall — Enforces deterministic security policies at the tool boundary for AI agents.
TraceRoot — Closes the self-improving loop for coding agents using production traces and evals.
OpenAPPA — Tracks data flows algebraically to prevent prompt injection without breaking agents.
abtop — Monitor multiple AI agent sessions and detect orphan ports from a single screen.
ARTEX — Runs autonomous multi-agent penetration testing with a dual-graph architecture.
iFixAi — Audits agent behavior independently without building manual eval harnesses.
Raindrop Workshop — Provides real-time observability and self-healing evaluation loops for local AI agents.
CyberStrike — Autonomous AI agents execute recon, web, network, and cloud penetration testing.
AiSOC — Runs locally without API keys with bundled models and auditable reasoning ledgers.
APIStack — Connect AI agents to over 100 tools using one API key or a single MCP endpoint.
VulnClaw — It uses LLM agents to autonomously execute end-to-end security testing loops.

All Numbat alternatives