dev-browser

<p align="center"> <img src="assets/header.png" alt="Dev Browser - Browser automation for Claude Code" width="100%"> </p>

Brought to you by Do Browser.

A browser automation tool that lets AI agents and developers control browsers with sandboxed JavaScript scripts.

Key features:

• Sandboxed execution - Scripts run in a QuickJS WASM sandbox with no host access
• Persistent pages - Navigate once, interact across multiple scripts
• Auto-connect - Connect to your running Chrome or launch a fresh Chromium
• Full Playwright API - goto, click, fill, locators, evaluate, screenshots, and more

Demo

https://github.com/user-attachments/assets/c6cf7fb9-b1dc-46ed-93b9-6e7240990c53

CLI Installation

npm install -g dev-browser
dev-browser install    # installs Playwright + Chromium

Quick start

# Launch a headless browser and run a script
dev-browser --headless <<'EOF'
const page = await browser.getPage("main");
await page.goto("https://example.com", { waitUntil: "domcontentloaded" });
console.log(await page.title());
EOF

# Connect to your running Chrome (enable at chrome://inspect/#remote-debugging)
dev-browser --connect <<'EOF'
const tabs = await browser.listPages();
console.log(JSON.stringify(tabs, null, 2));
EOF

PowerShell (Windows)

@"
const page = await browser.getPage("main");
await page.goto("https://example.com", { waitUntil: "domcontentloaded" });
console.log(await page.title());
"@ | dev-browser

With --connect:

@"
const page = await browser.getPage("main");
console.log(await page.title());
"@ | dev-browser --connect

Windows notes

PowerShell install:

npm install -g dev-browser
dev-browser install

To attach to a running Chrome instance on Windows:

chrome.exe --remote-debugging-port=9222
dev-browser --connect

Windows npm installs download the native dev-browser-windows-x64.exe release asset during postinstall, and the generated npm shims invoke that executable directly.

Using with AI agents

After installing, tell your agent to run dev-browser --help — the help output includes the current LLM usage guide and API reference.

For agents that discover local skills, install or refresh the embedded skill explicitly:

dev-browser install-skill --codex   # ~/.codex/skills/dev-browser/SKILL.md
dev-browser install-skill --claude  # ~/.claude/skills/dev-browser/SKILL.md
dev-browser install-skill --agents  # ~/.agents/skills/dev-browser/SKILL.md

Flags may be combined. With an interactive terminal, dev-browser install-skill prompts for targets. In non-interactive environments it updates all three locations, including Codex, so an older copied skill does not survive a CLI upgrade.

Idle browser cleanup

Daemon-launched named Chromium instances can be closed automatically after they have been idle for a configured duration:

dev-browser --idle-timeout 5m < script.js
DEV_BROWSER_IDLE_TIMEOUT_MS=300000 dev-browser status

The flag accepts 30s, 5m, 1h, or raw milliseconds. You can also set a user default in ~/.dev-browser/config.json:

{
  "idleTimeout": "5m"
}

Precedence is --idle-timeout, then DEV_BROWSER_IDLE_TIMEOUT_MS, then idleTimeout in the user config, then disabled. Set any source to 0 to disable cleanup. The effective setting is sent to an already-running daemon and shown by dev-browser status.

Cleanup is applied independently to each named browser. Activity is measured from both the start and completion of each request, so running requests are never reaped. Only Chromium instances launched by dev-browser are eligible; browsers attached with --connect are never closed by idle cleanup. Closing an idle browser does not delete its profile directory, cookies, or login state, and the next request relaunches it from the same persistent profile. dev-browser stop keeps its existing behavior of stopping the daemon and all managed browser connections.

<details> <summary>Allowing dev-browser in Claude Code without permission prompts</summary>

By default, Claude Code asks for approval each time it runs a bash command. You can pre-approve dev-browser so it runs without permission checks by adding it to the allow list in your settings.

Per-project — add to .claude/settings.json in your project root:

{
  "permissions": {
    "allow": [
      "Bash(dev-browser *)"
    ]
  }
}

Per-user (global) — add to ~/.claude/settings.json:

{
  "permissions": {
    "allow": [
      "Bash(dev-browser *)"
    ]
  }
}

The pattern Bash(dev-browser *) matches any command starting with dev-browser followed by arguments (e.g. dev-browser --headless, dev-browser --connect). This is safe because dev-browser scripts run in a sandboxed QuickJS WASM environment with no host filesystem or network access.

You can also allow related commands in the same list:

{
  "permissions": {
    "allow": [
      "Bash(dev-browser *)",
      "Bash(npx dev-browser *)"
    ]
  }
}

Tip: If you've already been prompted and clicked "Always allow", Claude Code adds the specific command pattern automatically. The settings file approach lets you pre-approve it before the first run.

</details> <details> <summary>Legacy Claude Code plugin installation</summary>

Claude Code

/plugin marketplace add sawyerhood/dev-browser
/plugin install dev-browser@sawyerhood/dev-browser

Restart Claude Code after installation.

</details>

Script API

Scripts run in a sandboxed QuickJS runtime (not Node.js). Available globals:

// Browser control
browser.getPage(nameOrId)    // Get/create named page, or connect to tab by targetId
browser.newPage()            // Create anonymous page (cleaned up after script)
browser.listPages()          // List all tabs: [{id, url, title, name}]
browser.closePage(name)      // Close a named page

// File I/O (restricted to ~/.dev-browser/tmp/)
await saveScreenshot(buf, name)   // Save screenshot buffer, returns path
await writeFile(name, data)       // Write file, returns path
await readFile(name)              // Read file, returns content

// Output
console.log/warn/error/info       // Routed to CLI stdout/stderr

Pages are full Playwright Page objects — goto, click, fill, locator, evaluate, screenshot, and everything else, including page.snapshotForAI({ track?, depth?, timeout? }), which returns { full, incremental? } for AI-friendly page snapshots.

Every page also exposes two computer-use toolsets:

• page.cua.* — pixel/vision tier: screenshot() saves a JPEG whose pixels map 1:1 onto CSS coordinates at any DPR and returns { path, width, height }; click, doubleClick, drag, move, scroll, keypress, and type act at those coordinates.
• page.domCua.* — DOM-id tier: getVisibleDom() snapshots visible interactive elements as pseudo-HTML lines with node_id=N; click, doubleClick, and scroll act by node id (ids are only valid against the latest snapshot of the current document), plus type and keypress for the focused element.

Benchmarks

| Method | Time | Cost | Turns | Success | | ----------------------- | ------- | ----- | ----- | ------- | | Dev Browser | 3m 53s | $0.88 | 29 | 100% | | Playwright MCP | 4m 31s | $1.45 | 51 | 100% | | Playwright Skill | 8m 07s | $1.45 | 38 | 67% | | Claude Chrome Extension | 12m 54s | $2.81 | 80 | 100% |

See dev-browser-eval for methodology.

License

MIT

Author

Sawyer Hood