Brig: microVM Sandbox for AI Coding Agents
Apache 2.0 open-source microVM that runs AI coding agents inside a secure, low-overhead isolation layer on Mac or Linux — hardware-enforced boundaries for Claude Code, Codex, Cursor, and more.
NOFire AI
Open-source
Free (Apache 2.0)
2026-09-16
Brig is an Apache 2.0 open-source microVM sandbox released by NOFire AI on September 15, 2026. It runs AI coding agents inside a secure, low-overhead isolation layer on macOS (Apple Silicon) or Linux (x86_64 and ARM), using hardware-level virtualization to enforce process boundaries physically rather than relying on OS-level permissions.
AI coding agents can install packages, execute arbitrary commands, use the network, and access developer credentials. Supply-chain attacks (like the Axios incident) become existential when an agent has full system access. The entire operating system is at risk during AI-accelerated development.
Brig closes this gap by containing the agent within the boundaries of an ephemeral microVM — enforced via hardware-level virtualization, not software process isolation.
- Hardware-enforced isolation: Process boundaries are physically enforced through hardware virtualization, not OS-level sandboxes
- Curated agent profiles: Pre-built profiles for Claude Code, Codex, Cursor, Gemini, Grok, and opencode
- Custom OCI images: Bring your own using OCI images (e.g., Ubuntu)
- Cross-platform: macOS on Apple Silicon, Linux on x86_64 and ARM
- Ephemeral sessions: Each agent run gets a fresh, disposable microVM
- Low audit surface: The microVMM code that a security-aware enterprise would audit is less than 20K lines — one of the lowest-overhead sandbox technologies available
- You run AI coding agents locally and want to prevent supply-chain attacks from reaching your host OS
- Your security team requires hardware-level isolation before approving agent use
- You need to run untrusted or experimental agent skills without risking your development machine
- You want to sandbox agent network access and credential exposure
- Minimal audit surface: <20K lines of code in the security-critical microVMM component
- No daemon overhead: Lightweight startup and teardown per agent session
- Familiar tooling: Curated profiles mean Claude Code, Cursor, etc. work without custom configuration
- Truly open: All components are Apache 2.0, free and open for audit by any security researcher
- macOS Apple Silicon only: No Intel Mac support; Linux requires x86_64 or ARM
- Resource overhead: microVMs add memory and CPU overhead vs. running agents directly, though Brig minimizes this
- Setup complexity: Initial configuration and OCI image preparation requires container knowledge
- New project: Released September 15, 2026 — production track record is still forming
Developers and security-conscious teams running AI coding agents on macOS or Linux who need hardware-enforced isolation without the overhead of a full virtual machine. Particularly valuable for teams where agents handle untrusted code, install third-party packages, or access sensitive credentials.