The AI agent sandbox in the browser tab
BrowserPod is a microVM that runs in the browser. Give agents a Linux-like environment on the user's device, with Node.js, Rust, git, bash and npm, and nothing to provision.
Why agents belong in the browser
Agents write code, install packages and run commands, and that work needs somewhere safe to happen, close to the person using your product.
BrowserPod moves that work into the user's tab. The agent's tools and output run inside the browser's sandbox, on the device of the person using your product, with the real Node.js, Rust, git and bash. Compute costs you nothing, and there's no pool to run out of.
Agents can adopt BrowserPod on their own, too. The library installs from npm and a pod boots without an account or a key, so an agent can add it to a project like any other dependency. When the work needs npm, GitHub or a model's API, BrowserPod Edge connects the pod.
What you need to know
Edge's default allowlist includes the major AI provider APIs, alongside the npm and Yarn registries and GitHub.
What people build
Coding agents in the browser
Claude Code, Codex CLI, Gemini CLI and Pi run unmodified in a tab, inside your product.
Run what the model writes
Execute generated scripts and programs on the user's device and stream the output back.
Agents that bring their own sandbox
An agent installs BrowserPod from npm and boots pods as it needs them, with no account.
Previews of agent-built apps
When the agent starts a dev server, Portals turn it into a URL the user can open.
What agents get
Everything an agent needs to write, install, build and run, inside a boundary built to contain it.
-
Sandbox
Agent output runs inside the browser's security boundary, away from the user's machine and your infrastructure.
-
Node.js, Rust, git, bash, npm
Claude Code, Codex CLI, Gemini CLI and Pi run unmodified, alongside the code they write.
-
Runs on the user's device
Every agent session runs on its user's device, so compute costs you zero.
-
No concurrency limits
A million users can each run their own agent, with no pool to exhaust.
-
Portals
When the agent starts a server, Portals give it a URL you can open and share.
From prompt to running code
-
Install
Add the npm package to your app, or let your agent install it. No key and no account.
-
Boot a pod
One call starts a Linux-like environment on the user's device.
-
Run what the agent writes
Write files into the pod and run them with node, bash or a Rust binary.
-
Connect with Edge
When the agent needs npm, GitHub or an AI provider's API, add BrowserPod Edge.
Run the agent's code
Write what the agent produced into the pod, then run it. The output streams to the terminal.
import { BrowserPod } from "@leaningtech/browserpod";
const pod = await BrowserPod.boot();
const terminal = await pod.createDefaultTerminal(document.querySelector("#console"));
const file = await pod.createFile("/home/user/task.js", "utf-8");
await file.write(agentOutput);
await file.close();
await pod.run("node", ["task.js"], { terminal, cwd: "/home/user" }); Questions
Can I run Claude Code in a pod?
Yes. Claude Code, Codex CLI, Gemini CLI and Pi run unmodified. Their API calls go through BrowserPod Edge, whose allowlist includes the major AI providers.
Can an agent use BrowserPod without an account?
Yes. The library is free for any use, and a pod boots without a key.
Can I see an agent running in a pod?
Yes. BrowserCode, our open playground, runs Claude Code, Codex CLI and Pi in the browser, with no server behind them.
What does it cost to run agents for a million users?
Nothing for compute, which runs on each user's device. Only pods connected through BrowserPod Edge are metered, per connected pod-hour.
Can we self-host BrowserPod?
Yes. Enterprise covers self-hosting the runtime and the network stack, vendoring it into your build and redistributing it in your product.










