perplexityai/numbat

Visibility into AI agent activity on endpoints, with on-device detection, optional pre-action blocking, and forensic reconstruction.

What it solves

numbat provides visibility and security monitoring for AI agents running on desktops, in CLIs, IDEs, and gateways. It addresses the lack of transparency in agent activity by detecting potentially malicious or unauthorized actions—such as accessing secret files or making unauthorized network requests—and providing the means to block those actions before they occur.

How it works

The tool uses local hooks, plugins, OTLP/HTTP logs, and on-disk session artifacts to observe agent behavior. It normalizes this activity into a single event model and evaluates it against a CEL (Common Expression Language) rule engine. Detection happens locally on the endpoint, and the system can either simply monitor activity or, if configured, use synchronous pre-action hooks to block specific actions based on defined rules.

Who it’s for

It is designed for security operators and developers who need to audit AI agent activity, perform forensic reconstruction of agent sessions, and enforce security policies on AI-powered tools within an organization.

Highlights

  • Live and Retroactive Monitoring: Supports real-time monitoring via hooks and plugins, as well as forensic reconstruction from on-disk artifacts without needing prior instrumentation.
  • Local Detection and Enforcement: Uses a local rule engine to detect threats and can optionally block actions via pre-action hooks.
  • Read-only Scanning: Includes tools for agent discovery and artifact scanning with built-in secret redaction.
  • Case Bundles: Allows the creation of portable, SHA-256 verified case bundles for investigations.
  • Cross-Platform: Distributed as a single binary for macOS, Linux, and Windows.

Related

  • Project
    eunomia-bpf/agentsight

    AgentSight is a kernel‑level observability tool for AI agents. Using eBPF and TLS tracing, it records prompts, model calls, subprocesses, file and network activity without requiring any SDK or proxy. It provides live dashboards (`top`), replay visualisations (`vis`), queryable SQLite logs, and OpenTelemetry export, helping developers debug failures, audit security, and understand token/resource costs of agents such as Claude Code, Gemini CLI, and Python‑based tools.

  • Project
    browser-act/skills

    A browser automation framework for AI agents that bypasses anti-bot protections and provides token-efficient web interaction via indexed text formats.

  • Project
    microsoft/agent-governance-toolkit

    Agent Governance Toolkit is a Microsoft‑maintained open‑source library that sits between autonomous AI agents and the tools they call, enforcing deterministic, auditable policies (allow/deny/require‑approval) written in YAML/OPA. It provides language‑specific SDKs (Python, TypeScript, .NET, Rust, Go), a CLI (`agt`) for compliance checks, and adapters for major agent frameworks, enabling production‑grade governance, attribution, and tamper‑evident logging.

  • Project
    openocta/openocta

    A desktop AIOps agent for Windows and macOS that automates IT operations tasks like alert analysis and system inspection through natural language and toolchain integration.

  • Project
    nightly-labs/openbot

    OpenBot is a local-first desktop app for managing persistent AI agents. It supports multiple AI providers, gives each agent its own workspace, and includes features like message queues, file transfers, an embedded browser, and agent-to-agent messaging. It's a development preview with full agent access, so caution is advised. Installation is available for macOS, Windows, and Linux, with options for servers, Docker, and Nix. The README also covers agent setup, development, marketplace, local skills, and importing agents from Grok Bot.