Smooth CLI – Give Your AI Agent a Browser That Works

TL;DR

TL;DR: Smooth CLI is a browser built for AI agents — it replaces messy click/type/scroll commands with a natural language interface, making agents 20x faster and 5x cheaper while working seamlessly with Claude Code and OpenClaw.

Source and Accuracy Notes

• Docs: docs.smooth.sh
• Playground: playground.smooth.sh
• HN: Show HN: Smooth CLI – Token-efficient browser for AI agents (109 points)

What Is Smooth CLI?

AI agents like Claude Code are incredibly powerful — but they’re mostly stuck in the terminal. Meanwhile, the most valuable work happens in the browser: research, form filling, QA testing, data extraction. Smooth CLI bridges this gap by giving your agent a real browser that speaks natural language.

Current browser tools for agents expose low-level actions: click, type, scroll. This creates three problems:

• Slow & expensive — Using a massive model to click buttons is wasteful. Every action costs tokens and time.
• Context pollution — Every keystroke fills up the context window with UI noise instead of your actual task.
• Wrong expertise — General-purpose models aren’t trained to handle iframes, shadow DOMs, and the messy reality of modern websites.

Smooth flips this around. Your agent says what it wants (“find all pricing plans on this page”, “fill out this contact form”). Smooth’s specialized model figures out the clicking.

Setup Workflow

Step 1: Install the CLI

# Install via npm
npm install -g @getsmooth/cli

# Or use curl for Linux/macOS
curl -fsSL https://get.smooth.sh/install.sh | sh

Step 2: Authenticate

smooth auth
# Opens browser for API key setup

Step 3: Configure Claude Code

# Point Claude Code to Smooth's browser
claude code --browser smooth

# Or set as default in ~/.claude/settings.json
{
  "browser": {
    "provider": "smooth",
    "api_key": "your-smooth-api-key"
  }
}

Step 4: Run Your First Task

# Natural language command
smooth browse "Find all pricing tiers on stripe.com and summarize them"

# Output: JSON with structured data
# No click coordinates, no waiting for selectors

Deeper Analysis

The Architecture

Smooth runs a lightweight browser in a managed environment with a small, fast model dedicated to DOM interaction. Your main agent (Claude Code, OpenClaw, etc.) sends high-level commands in plain English. Smooth’s execution layer translates those into clicks, scrolls, and form interactions — but abstracts away all the coordinate gymnastics.

The P2P tunnel is especially clever: traffic routes through your local machine, which means you can access localhost services, bypass CAPTCHAs that block cloud browsers, and reach geo-restricted content.

Performance Benchmarks

Smooth’s docs claim 20x speedup and 5x cost reduction compared to raw Playwright MCP or --chrome flags. While independent benchmarks aren’t public, the architecture makes sense: a fast, specialized model handling UI interactions while your main agent stays focused on reasoning.

Security Model

Smooth runs in an isolated environment with no permissions by default. You explicitly grant access per session. This is a meaningful improvement over giving an agent blanket browser access to your entire session.

Practical Evaluation Checklist

• Does it handle complex SPAs with shadow DOMs? Yes — specialized model trained on these
• Does it work with logged-in sessions? Yes — “Stay Logged In” feature persists cookies
• Can it upload/download files? Yes — drag-and-drop or path-based
• Does it support localhost? Yes — P2P tunnel routes through your machine
• Can you watch what it’s doing? Yes — “Watch Your Agent Work” mode streams the browser view

Security Notes

• Isolated execution environment by default
• Permission model: grant specific capabilities per session
• No persistent state between runs unless explicitly configured
• P2P tunnel routes traffic through your machine — review the tunnel settings before enabling

FAQ

Q: How does Smooth compare to Playwright MCP server?

A: Playwright MCP exposes low-level actions (click at X/Y, type into selector). Smooth takes natural language and hides the DOM complexity. For a coding agent, Smooth is dramatically more token-efficient — your agent says “submit the form” instead of “click button[data-testid=‘submit’], wait for networkidle, check response.”

Q: Does Smooth work with Cursor, Windsurf, or other AI IDEs?

A: Smooth has built-in integrations for Claude Code and OpenClaw. For other agents, check the “Other Agents” docs page — most agents that can invoke shell commands or use MCP can work with Smooth via its CLI interface.

Q: What’s the pricing model?

A: Free tier available with rate limits. Check smooth.sh/pricing for latest tiers. The CLI itself is open in that you can inspect what it’s doing — no black box.

Q: Can it handle CAPTCHAs?

A: The P2P tunnel routes traffic through your machine, which can help with some CAPTCHA implementations that block cloud IPs. However, this is not a guarantee — some CAPTCHAs are specifically designed to block automation.

Q: Does it work on Windows?

A: The primary CLI works on Linux and macOS. Windows support via WSL2 is documented. Native Windows is on the roadmap per community discussions.

Conclusion

Smooth CLI solves the browser gap that keeps stalling AI agent projects. Instead of burning tokens on UI navigation, your agent focuses on actual work. The natural language interface, P2P tunnel for localhost access, and session persistence features cover the most common pain points in agentic browser automation.

If you’re building anything that involves an AI agent interacting with the web — research, QA, data extraction, automated testing — Smooth is worth a look. The Claude Code integration alone justifies the setup time.

Reference: docs.smooth.sh/cli/overview