e2b
E2B Desktop Sandbox with autonomous computer use — cloud Linux environments for browser automation, visual UI grounding, screenshot analysis, and shell execution. Supports Browser Use cloud sessions,
SKILL.mdskills/e2b/SKILL.md
GitHubE2B — Cloud Desktop Sandbox & Computer Use
Secure cloud Linux desktops powered by E2B Desktop Sandbox with autonomous computer-use capabilities, Browser Use integration, and LLM-driven visual control.
Quick Start
# Set up environment
export E2B_API_KEY="your-e2b-api-key"
export BROWSERUSE_API_KEY="your-browseruse-api-key" # Optional: for cloud browser sessions
# Via daemon commands
/sandbox new # Spin up a code sandbox
/desktop new # Spin up a desktop sandbox (with display)
/desktop browse https://example.com # Open browser in desktop sandbox
/desktop screenshot # Capture current screen
/desktop click 512 384 # Click at coordinates
/desktop type "hello world" # Type text
/desktop key Return # Send key
/shell ls -la # Run shell command in sandbox
/sandbox kill # Terminate sandbox
Architecture
┌─────────────────────────────────────────────────────────────────┐
│ SolanaOS Daemon │
│ │
│ ┌──────────────────┐ ┌─────────────────┐ ┌───────────────┐ │
│ │ E2B Client │ │ Desktop Client │ │ Browser Agent │ │
│ │ (code sandbox) │ │ (desktop VNC) │ │ (autonomous) │ │
│ │ │ │ │ │ │ │
│ │ • RunCode() │ │ • Screenshot() │ │ • Vision LLM │ │
│ │ • RunCommand() │ │ • Click() │ │ • Action LLM │ │
│ │ • ListSandboxes │ │ • TypeText() │ │ • Tool Exec │ │
│ │ • KillSandbox() │ │ • SendKey() │ │ • OODA Loop │ │
│ │ │ │ • OpenBrowser() │ │ │ │
│ └──────┬───────────┘ │ • RunShell() │ └──────┬────────┘ │
│ │ │ • Scroll() │ │ │
│ │ │ • Stream() │ │ │
│ │ └────────┬─────────┘ │ │
│ │ │ │ │
│ ┌──────┴───────────────────────┴────────────────────┴────────┐ │
│ │ E2B REST API │ │
│ │ https://api.e2b.dev/sandboxes │ │
│ └────────────────────────────────────────────────────────────┘ │
│ │
│ ┌────────────────────────────────────────────────────────────┐ │
│ │ Browser Use Integration │ │
│ │ │ │
│ │ Cloud Providers: │ │
│ │ • Browser Use API (BROWSERUSE_API_KEY) │ │
│ │ • Browserbase (BROWSERBASE_API_KEY) │ │
│ │ │ │
│ │ Features: │ │
│ │ • CDP-based browser control │ │
│ │ • Cloud session management │ │
│ │ • Headed/headless modes │ │
│ │ • Profile persistence │ │
│ │ • Proxy support │ │
│ └────────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
Sandbox Types
Code Sandbox (Default)
Standard E2B code interpreter — Python execution, file I/O, shell commands. No display.
/sandbox new # Create code sandbox
/run print("hello world") # Execute Python
/shell pip install requests # Install packages
/sandbox kill # Terminate
Desktop Sandbox
Full Linux desktop with X11 display, VNC streaming, mouse/keyboard control.
/desktop new # Create desktop sandbox
/desktop screenshot # Capture screen
/desktop click 100 200 # Click at (100, 200)
/desktop type "search query" # Type text
/desktop key ctrl+a # Send key combo
/desktop browse https://... # Open browser
/desktop kill # Terminate
Browser Use Cloud Session
Cloud-hosted browser via Browser Use or Browserbase APIs.
/browseruse connect # Start cloud browser session
/browseruse open https://... # Navigate
/browseruse close # End session
Autonomous Computer Use (Browser Agent)
The Browser Agent combines vision LLMs with desktop control for autonomous task completion.
Agent Loop (OODA)
1. OBSERVE: Take screenshot of desktop
2. ORIENT: Send screenshot to vision LLM with objective
3. DECIDE: LLM analyzes screen, determines next action
4. ACT: Execute action (click, type, command, etc.)
5. REPEAT: Until objective is complete or max steps reached
Supported LLM Providers for Vision
The agent supports multiple vision LLM providers (configured via daemon):
| Provider | Models | Capabilities |
|---|---|---|
| OpenAI | GPT-4o, GPT-4o-mini | Vision + Action |
| Anthropic | Claude 3.5 Sonnet | Vision + Action |
| Gemini 2.0 Flash | Vision + Action | |
| Groq | Llama 3.2 | Vision + Action |
| OpenRouter | Qwen 2.5 VL, Llama 3.2 | Vision |
| Fireworks | Llama 3.2 | Vision |
| Mistral | Pixtral | Vision |
Agent Tools
| Tool | Description | Parameters |
|---|---|---|
click |
Left-click at coordinates | x, y |
double_click |
Double-click at coordinates | x, y |
right_click |
Right-click at coordinates | x, y |
type_text |
Type text into active input | text |
send_key |
Send key or combo | key (e.g. "Return", "ctrl+c") |
run_command |
Execute shell command | command |
open_browser |
Open browser to URL | url |
scroll |
Scroll mouse wheel | x, y, clicks |
screenshot |
Capture screen state | — |
stop |
Mark task as complete | — |
Daemon Integration
Chat Commands
| Command | Description |
|---|---|
/sandbox [new|kill|list] |
Manage code sandboxes |
/run <code> |
Execute Python in sandbox |
/shell <command> |
Execute shell in sandbox |
/desktop [new|kill|screenshot|click|type|key|browse] |
Desktop sandbox control |
/browseruse [connect|open|close|status] |
Browser Use sessions |
/agent browse <objective> |
Start autonomous browser agent |
Natural Language Detection
The daemon detects sandbox-related intents from natural language:
- "spin up a sandbox" →
/sandbox new - "open a browser and go to..." →
/desktop browse <url> - "take a screenshot" →
/desktop screenshot - "run this code" →
/run <code>
Playwright Integration
For structured E2E testing (vs. autonomous browsing), use the browse skill:
# Playwright tests run against the desktop sandbox
PLAYWRIGHT_BASE_URL="https://<sandbox-id>-8080.e2b.dev" npx playwright test
# Or against Browser Use cloud session
PLAYWRIGHT_CDP_URL="wss://..." npx playwright test --connect
Environment Variables
| Variable | Required | Description |
|---|---|---|
E2B_API_KEY |
Yes | E2B API key for sandbox creation |
BROWSERUSE_API_KEY |
No | Browser Use cloud API key |
BROWSERBASE_API_KEY |
No | Browserbase API key (alternative) |
BROWSERBASE_PROJECT_ID |
No | Browserbase project ID |
ANTHROPIC_API_KEY |
No | For vision/action LLM (Claude) |
OPENAI_API_KEY |
No | For vision/action LLM (GPT-4o) |
GROQ_API_KEY |
No | For vision/action LLM (Llama) |
OPENROUTER_API_KEY |
No | For vision LLM (multi-model) |
API Reference
E2B REST API
POST /sandboxes Create sandbox
GET /sandboxes List sandboxes
DELETE /sandboxes/:id Kill sandbox
POST /sandboxes/:id/code Execute code
POST /sandboxes/:id/commands Run shell command
GET /sandboxes/:id/screenshot Capture desktop (desktop template)
POST /sandboxes/:id/timeout Extend timeout
Browser Use API
POST https://api.browser-use.com/api/v2/browsers Create session
PATCH https://api.browser-use.com/api/v2/browsers/:id Stop session
Browserbase API
POST https://api.browserbase.com/v1/sessions Create session
POST https://api.browserbase.com/v1/sessions/:id Close session