E2B Desktop — Headless Linux Desktop for AI Computer Use

Quick Use

1. pip install e2b-desktop
2. Sandbox.create() returns a desktop with desktop.stream.get_url() for VNC
3. Drive with desktop.left_click(x, y), desktop.write(...), desktop.screenshot()

Intro

E2B Desktop is a managed cloud Linux desktop — full X11 environment, Firefox, terminal, file manager — exposed via VNC and an SDK. Drop in as the target environment for Claude Computer Use or OpenAI's CUA without setting up Docker / X11 locally. Best for: production deployments of computer-using AI agents, multi-tenant SaaS where each user gets an isolated desktop. Works with: Python SDK, Claude Computer Use, any agent that drives a screen+keyboard+mouse. Setup time: 5 minutes.

Spawn a desktop sandbox

from e2b_desktop import Sandbox

with Sandbox.create() as desktop:
    # Get the live VNC URL (embed in iframe or watch via VNC viewer)
    print(desktop.stream.get_url())

    # Take a screenshot
    screenshot = desktop.screenshot()
    with open("screenshot.png", "wb") as f:
        f.write(screenshot)

    # Drive mouse + keyboard
    desktop.left_click(500, 300)
    desktop.write("hello world")
    desktop.press("Enter")

    # Launch an app
    desktop.launch("firefox", "https://news.ycombinator.com")

Use with Claude Computer Use

from anthropic import Anthropic
from e2b_desktop import Sandbox

desktop = Sandbox.create()

response = anthropic.beta.messages.create(
    model="claude-3-5-sonnet-20241022",
    max_tokens=4096,
    tools=[
        {
            "type": "computer_20241022",
            "name": "computer",
            "display_width_px": desktop.width,
            "display_height_px": desktop.height,
        }
    ],
    messages=[{"role": "user", "content": "Find the top story on HN, screenshot it"}],
    extra_headers={"anthropic-beta": "computer-use-2024-10-22"},
)

# Each tool call from Claude — execute against desktop:
for content in response.content:
    if content.type == "tool_use" and content.input.get("action") == "screenshot":
        ss = desktop.screenshot()
        # ... return base64 to Claude
    elif content.input.get("action") == "left_click":
        x, y = content.input["coordinate"]
        desktop.left_click(x, y)
    # ... handle other actions

Multi-tenant pattern

# Each user gets their own sandbox
def get_user_desktop(user_id):
    sandbox = Sandbox.create(timeout=3600)
    sandbox.set_metadata("user_id", user_id)
    return sandbox

# Reconnect to a running sandbox by ID
sandbox = Sandbox.connect(sandbox_id)

Sessions persist across calls — useful for "agent did some work, returned to the user, now resumes" flows.

FAQ

Q: How is this different from Open Interpreter OS Mode? A: Open Interpreter OS Mode runs on YOUR machine — your screen, your keyboard. E2B Desktop runs in the cloud — agent gets an isolated cloud desktop, your machine is uninvolved. Use OS Mode for personal automation; E2B Desktop for production / multi-tenant SaaS.

Q: Can I install custom apps in the desktop? A: Yes — E2B Desktops use Ubuntu under the hood. Run desktop.commands.run('apt install --yes <package>') or build a custom image with e2b template build for fast cold starts.

Q: Does it support audio / video? A: VNC streams the visual desktop in real time. Audio is not yet supported; video apps work for visual-only output. For audio interactions, layer a separate audio pipeline.

Source & Thanks

Built by E2B. Licensed under Apache-2.0.

e2b-dev/desktop — ⭐ 1,500+