Numbat

Numbat is an open-source security enforcement harness and policy engine developed by Perplexity AI. Written in Go, it is engineered to inspect, intercept, and govern actions executed by AI coding agents and development environments. In day-to-day operation, Numbat functions as an execution guardrail that evaluates agent tool calls and commands against deterministic security policies. It evaluates rules using embedded Common Expression Language (CEL) expressions to validate shell executions, PowerShell commands, and file operations before they are executed in the development environment. What distinguishes Numbat from conventional agent orchestrators is its strict focus on security policy enforcement and tamper prevention. Rather than focusing on prompt routing or agent autonomy, it provides runtime enforcement hooks (including integrations for editors like Cursor) to prevent autonomous agents from modifying critical agent configurations or running unauthorized scripts. Numbat is open-source software maintained by Perplexity AI under a standard GitHub open-source repository structure. It operates locally as a compiled Go binary or runtime hook without mandatory cloud platform subscriptions.

Tags: AI, ai agent, cli, developer tools, OpenSource

Visit Numbat

Key features

  • CEL-Based Policy Engine — Evaluates incoming commands and agent actions against embedded Common Expression Language expressions. - Agent Configuration Guardrails — Enforces policies designed to prevent autonomous agents from tampering with system and agent configuration files. - Shell and PowerShell Projections — Inspects and hardens call targets across varied shell environments, including PowerShell call projections. - Editor and Agent Hooks — Integrates into AI coding agent workflows, including hook-based enforcement for editors such as Cursor. - Compiled Go CLI Binary — Delivers fast evaluation and low runtime overhead directly within local development workflows.

Use cases

Use Case: Preventing Agent Configuration Tampering When running autonomous coding agents with repository write permissions, Numbat intercepts file modification calls to block unauthorized edits to agent definitions, system prompts, or security configurations.

Use Case: Restricting Malicious or Malformed Shell Commands During automated code execution loops, Numbat inspects shell and PowerShell invocations against CEL policies to catch and reject empty targets or risky execution patterns before subprocesses spawn.

Use Case: IDE Hook Enforcement in Agentic Workflows Teams deploying IDE-based agents like Cursor can configure Numbat hooks to enforce consistent security and policy boundaries across all developer machines.

Who it is for

Numbat is built for platform engineers, security teams, and developers running autonomous AI coding agents who need deterministic runtime guardrails and execution policies. It is less suitable for users looking for an interactive conversational UI or an out-of-the-box coding assistant by itself.

Install guide

Installation Clone the repository and build the Go binary: bash git clone https://github.com/perplexityai/numbat.git cd numbat go build -o numbat ./cmd/numbat

First Run Verify the compiled binary and check available runtime commands: bash ./numbat --help

Example Session Run Numbat to evaluate security rules and hook into your agent's command execution pipeline: bash # Validate local rule definitions ./numbat check ./rules # Run enforcement harness ./numbat enforce --config ./rules/default.cel

Part of CLI Coding Agents

Categories: AI Coding Agents, Developer Tools

Related CLI agents

  • Skillreaper — Skillreaper is an open-source skill lifecycle management and auditing tool for AI agent ecosystems, developed by thousandflowers. It analyzes agent…
  • ActPlane — ActPlane is an open-source agent execution and policy enforcement harness developed by the Eunomia-bpf organization. It functions as a security…
  • Zosma QA — Zosma QA is an open-source testing-focused agent harness and orchestration framework created by Zosma AI. It provides scaffolding, structured prompt…
  • Stratless — Stratless is an open-source agent harness and execution orchestrator developed by the Stratless AI team. It provides a modular infrastructure…
  • Agent Island — Agent Island is an open-source agent orchestration harness created by developer tristan666666 on GitHub. Designed as a framework for running…
  • Repo Forensics — Repo Forensics is an open-source agent orchestration harness and repository security gating tool developed by Alex Greenshpan (alexgreensh). Designed to…
  • AI Tools
  • Categories
  • Industries
  • CLI Coding Agents
  • MCP Servers
  • MCP Categories