A faster, lighter, cheaper alternative to sandboxes — with agent orchestration built in.
import { AgentOs } from "@agent-os/core";
const vm = await AgentOs.create();
// Create an agent session
const { sessionId } = await vm.createSession("pi");
// Stream events (tool calls, text output, etc.)
vm.onSessionEvent(sessionId, (event) => console.log(event));
// Send a prompt and wait for the response
await vm.prompt(sessionId, "Write a Python script that calculates pi");
Pi
Claude Code
Codex
OpenCode
AmpA new architecture.
Built from the ground up for lightweight agents. agentOS provides the flexibility of Linux with lower overhead than sandboxes.
WebAssembly + V8 Isolates
High-performance virtualization without specialized infrastructure. The same battle-hardened isolation technology that powers Google Chrome.
Battle-tested technology
You're probably using this technology right now to view this page. Bring the same power to your agents. No VMs, no containers, no overhead.
Performance benchmarks
agentOS vs. traditional sandboxes.
Workload: Pi coding agent session with MCP servers and mounted file systems
Why the gap: agentOS runs agents in-process — V8 isolates and Wasm inside your host. No VM to boot, no network hop, no disk image. Sandboxes must boot an entire environment, allocate memory, and establish a network connection before code can run.
Sandbox baseline: E2B, the fastest mainstream sandbox provider as of March 30, 2026.
agentOS: Median of 10,000 runs (100 iterations x 100 samples) on Intel i7-12700KF.6.1 ms
Why the gap: In-process isolates share the host's memory. Each additional execution only adds its own heap and stack. Sandboxes allocate a dedicated environment with a minimum memory reservation, even if the code inside uses far less.
Sandbox baseline: Daytona, the cheapest mainstream sandbox provider as of March 30, 2026. Default sandbox: 1 vCPU + 1 GiB RAM.
agentOS: ~131 MB for a full Pi coding agent session with MCP servers and file system mounts.~131 MB
Sandboxes reserve idle RAM per agent.
server price per second / concurrent executions per serverWhy it's cheaper: Each execution uses ~131 MB instead of a ~1024 MB sandbox minimum. And you run on your own hardware, which is significantly cheaper than per-second sandbox billing.
Sandbox baseline: Daytona, the cheapest mainstream sandbox provider as of March 30, 2026. Default sandbox: 1 vCPU + 1 GiB RAM at $0.0504/vCPU-h + $0.0162/GiB-h.
agentOS: ~131 MB baseline per execution, assuming 70% utilization (industry-standard HPA scaling threshold). Select a hardware tier above to compare.$0.00000058/s
Assumes one agent per sandbox, needed for isolation.
Measured on Intel i7-12700KF. Cold start baseline: E2B, the fastest mainstream sandbox provider as of March 30, 2026. Cost baseline: Daytona, the cheapest mainstream sandbox provider as of March 30, 2026 (1 vCPU + 1 GiB default). Cost assumes 70% utilization on self-hosted hardware vs. per-second sandbox billing. Benchmark document
Meet your agent's new operating system.
Any agent.
Any supported harness, including Claude Code, Codex, and Pi, runs inside the OS instead of beside it.
The agent speaks ACP. agentOS brokers every tool call, file read, and network request it makes, then streams every response back. The host stays in control of what the agent can see and do.
Tools and resources over MCP.
Expose any API, toolchain, or data source as a tool or resource the agent can use.
Host-defined tools and MCP servers plug straight into the session. The agent calls them like local functions while agentOS handles routing and scoping, so it never holds a raw API key.
Durable agent sessions.
Every run is a managed session with its own state, history, and lifecycle.
Start, pause, resume, and replay. Session state lives in the host, so you can attach lifecycle hooks, persist transcripts, and pick a conversation back up exactly where it left off.
Extend with a sandbox.
agentOS runs most work in-process and reaches for a full sandbox when a job needs more.
Hand heavier or untrusted workloads, like a real kernel, native binaries, or a GPU, to a sandbox without leaving the session. agentOS handles the fast path while sandboxes cover the long tail.
Orchestrate agents and work.
Let agents delegate to other agents through host-defined tools, workflows, and cron.
Fan work out to sub-agents, coordinate multi-step flows, and schedule recurring jobs. The host brokers every delegated call so it runs under the same limits and permissions.
agentOSEverything agents need, built in.
Isolation, resource limits, networking, and storage in one npm package, no glue code.
Run any agent harness
Claude Code, Codex, OpenCode, Amp, and Pi all run behind one unified API. Swap or add agents without touching your infrastructure.
One agent can't crash the rest
Each agent runs in its own V8 isolate with a bounded heap, stack, and CPU slice. A memory bomb or infinite loop ends that isolate while the host and every other agent keep running.
Restrict CPU & memory granularly
Set precise resource limits per agent. No runaway processes, no noisy neighbors.
Programmatic network control
Allow, deny, or proxy any outbound connection. Full control over what your agents can reach.
Mount anything as a file system
S3, GitHub, databases. No per-agent credentials needed, since the host handles access scoping.
Embed in your backend
agentOS is a library, not a service. Drop it into your existing API and toolchain, and agents call your code as plain JavaScript functions or hooks with no separate agent auth to wire up.
Deploy anywhere
Just an npm package, with no vendor lock-in and no special infrastructure. Run agents on your laptop, Railway, Vercel, Kubernetes, ECS, Lambda, or on-prem. Wherever your code runs, agents run.
agentOS Registry
A marketplace for agent capabilities. Browse and install pre-built tools, integrations, file systems, databases, and sandboxes — one command away.
