rivet-dev/agentos

What it solves

agentOS provides a high-performance, secure, and lightweight execution environment for AI agents. It replaces heavy, slow-to-boot Linux sandboxes (microVMs or containers) with a library-based virtual machine that runs directly inside a backend process, drastically reducing cold start times, memory usage, and operational costs.

How it works

It runs agents in fully virtualized VMs where a trusted sidecar process manages the kernel (filesystem, process table, network stack) and brokers all guest syscalls. Guest JavaScript is executed via native V8 isolates and compiled tools run as WebAssembly. This architecture allows multiple VMs to share a single sidecar process, avoiding the overhead of full OS processes. It also supports "sandbox mounting," allowing it to connect to full Linux sandboxes for heavy workloads like native compilation.

Who it’s for

Developers building AI agent backends who need to give agents the ability to execute code, run shell scripts, and interact with filesystems or networks without the latency and cost of traditional cloud sandboxes.

Highlights

• Extreme Performance: Cold starts in single-digit milliseconds and significantly lower memory footprints compared to traditional sandboxes.
• Deep Integration: Agents can call host functions directly via bindings as ordinary JavaScript calls rather than network services.
• Granular Security: Default-deny network egress and strict permissions for filesystem, process, and environment access.
• Built-in Agent Support: Native support for agents like Pi, Claude Code, Codex, and OpenCode via a unified API.
• Extensible Registry: An npm-based registry for adding POSIX utilities, command packages (e.g., git, sqlite3), and cloud browser integrations.