Smooth CLI – Token-Efficient Browser for AI Agents
Smooth CLI gives AI coding agents a goal-oriented browser that navigates the web via natural language instead of DOM clicks, cutting cost and latency dramatically.
TL;DR
TL;DR: Smooth CLI replaces low-level browser automation with a goal-oriented interface — tell it what to accomplish in plain language, and it handles the DOM so your agent doesn’t have to.
Source and Accuracy Notes
⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.
- Project page: smooth.sh — verified
- Source repository: github.com/circlemind-ai/smooth-sdk — verified (README read in full)
- License: Not specified in repository (pyproject.toml has no license field) — verified 2026-08-07
- HN launch thread: news.ycombinator.com/item?id=42582209 — 109 points, 2026-02-05
What Is Smooth CLI?
Most AI coding agents — Claude Code, Cursor, Cline, Windsurf, and similar — run entirely in the terminal. Meanwhile, the most valuable human workflows happen in the browser. Bridging that gap usually means Playwright MCP or Chrome flags that expose hundreds of low-level actions (click(x=342, y=128), type(...), scroll(...)). These create three problems:
- Cost and latency — routing every button click through a frontier-sized model is expensive
- Context pollution — every DOM interaction chews through the context window
- Lack of navigation expertise — general models are not web automation specialists
Smooth CLI flips the model. Instead of micromanaging the browser, you give the agent a goal in natural language and Smooth handles the mechanics. The agent focuses on what to accomplish; Smooth handles how to click, scroll, and extract.
# Instead of dozens of low-level actions:
click(x=342, y=128)
type("flights from NYC to LA")
click(x=401, y=130)
scroll(down=500)
# ... 50 more steps
# You just say:
smooth run <session-id> "Find the cheapest flight from NYC to LA today"
Setup Workflow
Step 1: Install
pip install smooth-py
Requires Python 3.10 or later.
Step 2: Configure API Key
smooth config --api-key <your-key>
Get your API key at app.smooth.sh. The CLI also has a free tier.
Step 3: Install the Claude Code Skill (optional)
For Claude Code integration:
npx skills add https://github.com/circlemind-ai/smooth-sdk
Step 4: Run a Session
# Start a browser session
smooth start-session
# Run a task on a specific URL
smooth run <session-id> "Find the pricing page and extract all plan names and prices" --url "https://example.com"
# Close when done
smooth close-session <session-id>
The session gives you a live URL you can inspect in a regular browser while the agent works. Multiple sessions can run concurrently for parallel tasks.
Deeper Analysis
Why It Matters for Agent Workflows
The Playwright MCP approach treats the browser as a universal interface that every agent should learn to operate. Smooth’s bet is the opposite: the browser is a means to an end, and agents shouldn’t need to be browser experts. By abstracting the DOM behind goal-oriented commands, Smooth lets models concentrate reasoning budget on actual tasks rather than mechanical execution.
Performance Claims
The README claims Smooth is 20x faster and 5x cheaper than Claude Code with --chrome. These figures come from the project’s own benchmarking (circlemind-ai/smooth-sdk README). They reflect internal tests against Playwright-based browser tools, not independent third-party benchmarks. Treat the numbers as directional rather than guaranteed.
Compatibility
The README explicitly lists support for:
- Claude Code
- Cursor
- Cline
- Codex
- Windsurf
- OpenCode
- Kiro
- Factory AI
- Github Copilot
- Antigravity
- Clawdbot / Moltbot / OpenClaw
Beyond that, any agent that can execute CLI commands can use Smooth.
Pricing
Based on the docs.smooth.sh/pricing page (scraped 2026-08-07), the pricing ladder includes a free tier plus paid plans at $5, $50, and $500. The exact tier boundaries are not reproduced here since pricing pages change frequently — check app.smooth.sh for current rates.
Practical Evaluation Checklist
- [ ]
pip install smooth-pycompletes without error on Python 3.10+ - [ ]
smooth config --api-keyaccepts a key without error - [ ]
smooth start-sessionopens a new browser session and returns a session ID - [ ]
smooth runaccepts a natural-language goal and returns a live URL - [ ] Session live URL is accessible in a regular browser
- [ ] Multiple concurrent sessions work independently
- [ ]
smooth close-sessionterminates the browser cleanly
Security Notes
- API key is stored locally via
smooth config— keep the config file permissions restricted - The CLI uses your machine’s IP address, which helps avoid captchas on sensitive sites
- No open-source browser engine is bundled; Smooth manages its own browser runtime
FAQ
Q: Does Smooth work with agents other than Claude Code? A: Yes. Any agent that can invoke CLI commands can use Smooth, including Cursor, Cline, Windsurf, Codex, and others listed in the README.
Q: Is the browser visible while it runs? A: Yes. Each session exposes a live URL so you can watch the browser execute tasks in real time.
Q: What happens if a task requires logging in? A: Smooth sessions run on your machine with your IP, but authentication is between you and the target site. The README does not document a session-sharing or cookie-import mechanism.
Q: Is this open source? A: The SDK lives at github.com/circlemind-ai/smooth-sdk and is publicly accessible, but the license field in pyproject.toml is empty. Verify the license directly from the repository before assuming redistribution rights.
Conclusion
Smooth CLI solves the browser-automation problem for AI agents by inverting the abstraction level. Rather than exposing DOM primitives, it accepts goal-oriented instructions and manages the browser internally. For workflows that require the terminal to interact with the web — price research, data extraction, automated form filling — this is a meaningful reduction in friction.
The free tier at app.smooth.sh lets you evaluate whether the performance and cost claims hold for your specific use case before committing to a paid plan.
Related Posts
ai-setup
Recall – Persistent Memory for Claude Code via MCP Hooks
Recall gives Claude Code a permanent memory store that survives session restarts and context compaction. Four hooks capture and restore context automatically — with cloud SaaS or self-hosted options.
2/28/2026
dev-tools
Automotive Skills Suite for AI Engineering
Evaluate Automotive Skills Suite for APQP, ASPICE, HARA, safety-plan, and DIA workflows with setup notes, governance risks, and SME review guidance.
5/28/2026
dev-tools
awesome-agentic-ai-zh Roadmap Guide
Explore awesome-agentic-ai-zh as a Chinese agentic AI learning roadmap, with setup notes, track selection, study workflow, and evaluation guidance.
5/28/2026