Semantic tools, a locked-down browser, and a human one click away.

Run Chromium in Docker, give the agent semantic tools instead of raw scripts, restrict where it can navigate, and let a person take over through a password-protected desktop that shares the same browser.

From container to agent action.

The browser runs in your Docker, under your policy, and is destroyed when the run ends.

  1. 01

    Create the browser

    A DockerBrowserClient backed by a sandbox client starts Chromium with your workspace, network, and resources.

  2. 02

    Wait until ready

    Wait for the automation and desktop capabilities before connecting.

  3. 03

    Expose semantic tools

    createBrowserTools() picks the tools and the navigation policy for this connection.

  4. 04

    Run the agent

    Pass the tools to an Agent. It reads a snapshot before each action.

  5. 05

    Clean up

    Disconnect and destroy the browser, or stop it and resume it later.

Automation with the safety on.

Semantic tools

Tabs, navigate, ARIA snapshot, click, type, press key, and screenshot. No arbitrary scripts or raw CDP.

Navigation policy

Lock navigation to exact origins. Private and reserved IP literals are rejected on the first URL and every redirect.

Human takeover

A noVNC desktop shares the agent’s Chromium. A renewable lease pauses new agent actions while you drive.

Hardened container

Non-root Chromium with its sandbox on, seccomp, and dropped capabilities. It never falls back to --no-sandbox.

Give an agent a locked-down browser.

Choose the tools and the origins. Everything else is denied.

browser-agent.ts
1import { Agent } from '@anvia/core/agent'2import { DockerBrowserClient, createBrowserTools } from '@anvia/browser'3import { DockerSandboxClient } from '@anvia/sandbox'45const browserClient = new DockerBrowserClient({6  sandboxClient: new DockerSandboxClient(),7  image: process.env.ANVIA_BROWSER_IMAGE!,8})9const browser = await browserClient.createBrowser({10  workspace: { type: 'ephemeral' },11  network: { mode: 'bridge' },12})13await browser.waitForCapabilities({ capabilities: ['automation', 'desktop'], timeoutMs: 30_000 })1415const connection = await browser.connect({ timeoutMs: 30_000 })16const tools = createBrowserTools({17  connection,18  tools: ['browser_navigate', 'browser_snapshot', 'browser_click', 'browser_type'],19  navigation: { mode: 'origins', origins: ['https://app.example.com'] },20})21const agent = new Agent({22  id: 'browser-agent',23  model,24  instructions: 'Inspect a snapshot before every browser action.',25  tools,26})

Human takeover does not authenticate users or authorize product operations. The navigation policy does not block third-party subresources, and Docker bridge networking is not SSRF isolation. Requires Node 20.12+ and Docker.

Build it.

Start from the guide, run the example, or look at what else Anvia handles.

Back to the SDK

Anvia
GitHubGet started