Browser MCP is one path. For most coding agents, the shorter line is a CLI and an Agent Skill.
Notte gives your AI agent a real browser in the cloud - stealth fingerprints, residential proxies, CAPTCHA solving, persistent session profiles, and parallel sessions. We expose that browser through three integration surfaces: the Notte CLI, the notte-browser Agent Skill for 30+ coding agents, and the Notte Browser MCP server for desktop MCP clients. Same cloud browser, same API key, three different shapes for three different kinds of agent.
"Browser MCP" has become a catch-all term for any way of giving an AI agent a browser. In practice there are three integration shapes a Notte user actually picks between, and the right one depends on what their agent looks like rather than which protocol is trendy this quarter.
For coding agents - the Claude Codes, Cursor agent modes, Goose, OpenHands, Gemini CLI, OpenAI Codex - the browser arrives through the Notte CLI, and the agent learns to drive it through the official notte-browser Agent Skill. For desktop MCP-native clients like Claude Desktop and Windsurf, the same browser is exposed through the Notte Browser MCP server. All three resolve to the same Notte account, the same session profiles, the same vaults, the same Functions.
The rest of this page walks through each surface, what it is good for, and how it compares to the other "Browser MCP" options in the market. The order is intentional: CLI first, Skill second, MCP last. That mirrors how most teams ship today.
The universal substrate for agents that drive a terminal.
Every modern coding agent already knows how to run shell commands. That is the integration surface they were trained on and the one their host environments give them by default. A well-designed CLI is therefore the shortest line between an agent and a browser - no MCP transport to negotiate, no JSON config to maintain across machines, no permission dance that changes with every client release.
The Notte CLI is a single Go binary that exposes every concept behind notte.cc as a resource-based command. Browser sessions, page actions, AI agents, scraping, personas, secret vaults, profiles, and scheduled Functions all live behind one verb tree that an agent can discover with notte --help. API keys live in the system keychain. Output flips between human-readable tables and clean JSON for piping. The whole thing is built so that a coding agent can pick it up cold and still produce a working invocation on the first try.
Concretely: an agent can start a stealth session, navigate to a URL, observe interactive elements, click and fill them, scrape a structured payload, and tear the session down in seven commands - all from inside whatever shell its IDE happens to hand it. No client-specific JSON, no protocol mismatch when the host editor upgrades, no debugging an MCP transport that went silent.
The CLI ships through Homebrew for one-command install, with go install and source builds as fallbacks. Authentication is a single notte auth login that drops a key into the OS keychain on macOS, Linux, and Windows. Shell completions ship for bash, zsh, fish, and PowerShell. The full command reference and the resource-by- resource walkthrough live in the repo README.
How coding agents actually learn to drive the Notte CLI.
A CLI is necessary but not sufficient. Without context, a fresh coding agent staring at a new tool will guess at flags, miss the resource-based command structure, and end up burning tokens on --help calls. Agent Skills are the convention that closes that gap: a small bundle of markdown and metadata that an agent loads on demand when it encounters a task that matches the skill's description.
The notte-browser skill is the official one for Notte. It ships the observe/click/fill/scrape primitives, the session lifecycle, the auth flow, the profile model, the Functions deploy story, and the common gotchas - all in a form the agent's host runtime can pull in only when the task actually calls for a browser. The skill is open source, MIT-licensed, and lives in the same repo as the marketplace metadata.
The reach is broader than Claude. Agent Skills is an open packaging spec, and the notte-browser skill is consumed today by Claude Code, Cursor, Goose, OpenHands, Gemini CLI, OpenAI Codex, Kiro, VS Code, Bitterbot, Windsurf, Cline, Continue, and 20+ other skill-aware coding agents. Install once, and the same body of context lights up every editor your team is in.
Installation matches whatever your host expects. Claude Code users add the marketplace and install the plugin in two chat-bar commands. Cursor users install from the official Cursor Plugin Marketplace. Any other skill-aware agent picks the skill up via the skills tool. The prerequisite in every case is the Notte CLI from section 01 plus an API key.
For desktop MCP clients that prefer tools over shell commands.
Not every AI client is a coding agent in a terminal. Claude Desktop and a handful of MCP-native editors expect their tools to arrive over the Model Context Protocol - a JSON-RPC layer defined by Anthropic for connecting assistants to external capabilities. For those clients, Notte ships an MCP server that wraps the same cloud browser used by the CLI and the Skill.
The trade-off is honest: MCP gives you nicer in-client tool previews and a permission model that some teams prefer; the CLI gives you universal reach across coding agents and one less moving part. If you are reaching for a browser inside Claude Desktop, the MCP path is the right one. If you are reaching for a browser inside Claude Code, Cursor agent mode, or any other agent that runs commands, the CLI plus Skill is shorter and more portable.
Setup, current transport options (stdio, SSE), the tool schema, and the up-to-date status of hosted vs self-hosted endpoints live in the docs. Treat that page as the source of truth - the MCP surface area moves faster than this page does.
Pick the CLI
Shell scripts, CI/CD, scheduled jobs, manual one-offs, and any coding agent that drives a terminal. Start here if you are not sure.
Pick the Skill
Claude Code, Cursor, Goose, OpenHands, Gemini CLI, OpenAI Codex, and every other Agent-Skills-compatible coding agent where you want the model to know how to use the CLI automatically.
Pick the MCP
Claude Desktop, Windsurf, and other MCP-native desktop clients where the integration surface is the Model Context Protocol rather than a shell.
For the keyword shopper: this is the honest table. Playwright MCP and @browsermcp/mcp drive a local browser; they are free and fine for simple, public, low-volume work. Browserbase MCP runs in the cloud with stealth but ships no agent loop and no CLI/Skill surface. Notte is the option when you want stealth, persistence, parallel cloud sessions, and reach across every coding-agent ecosystem in one product.
| Feature | Notte | Playwright MCP | @browsermcp/mcp | Browserbase MCP |
|---|---|---|---|---|
| Runs in the cloud | Yes | No (local only) | No (local Chrome) | Yes |
| Stealth + proxies + CAPTCHA | Built-in | No | No | Add-on |
| Works in a terminal-driven agent | Yes (CLI + Skill) | Workarounds | No | Workarounds |
| Works in MCP desktop clients | Yes | Yes | Yes | Yes |
| Persistent sessions | Yes (profiles + vaults) | Manual | Per-tab | Yes |
| Built-in AI agent loop | Yes (notte agent) | No | No | No |
| Open source | Yes (CLI, Skills, SDK) | Yes | Yes | Partly |
All three surfaces back onto the same cloud browser - the one Notte ships in production and the one that currently sits at the top of Browser Arena, the open benchmark for cloud browser providers. Stealth fingerprints, residential proxies, CAPTCHA solving, persistent session profiles, and parallel sessions are not add-ons on top of the integration; they are the floor.
That is the point of having three surfaces in the first place. The browser - and the work that goes into keeping it fast, cheap, and undetectable - is the asset. Whether you drive it from a CLI, a Skill, or an MCP server is a question about your agent's runtime, not about which Notte product you are buying.
What is Browser MCP?
Browser MCP is the general name for any server that follows the Model Context Protocol and exposes a web browser as a set of tools an AI assistant can call. Notte ships one of those, but for coding agents that live in a terminal we recommend the Notte CLI driven by our official Agent Skill instead - it skips the MCP transport layer entirely and lets the agent drive the same Notte cloud browser through shell commands it already knows how to run.
Why a CLI instead of an MCP server for coding agents?
Every modern coding agent - Claude Code, Cursor agent mode, Cline, Continue, Goose, OpenHands, Gemini CLI, OpenAI Codex - already runs shell commands. A CLI is the universal surface. There is no protocol to negotiate, no JSON config to maintain across machines, and the agent gets full discoverability via `notte --help`. The MCP path makes sense for MCP-native desktop clients (Claude Desktop, Windsurf); for coding agents, the CLI is the shorter line.
What is the Notte Agent Skill?
The notte-browser Agent Skill (github.com/nottelabs/notte-skills) is an open-source skill bundle that teaches any Agent-Skills-compatible coding agent how to drive the Notte CLI. Install it once with `npx skills add nottelabs/notte-skills` (or via the Claude Code plugin marketplace) and the agent learns the observe/click/fill/scrape primitives, session lifecycle, and the Notte Functions deploy flow without any further configuration.
Which clients does the Agent Skill support?
Claude Code, Cursor, Goose, OpenHands, Gemini CLI, OpenAI Codex, Kiro, VS Code, Bitterbot, Windsurf, Cline, Continue, and 20+ other agents that implement the Agent Skills spec. Cursor users can also install the skill through the Cursor Plugin Marketplace.
Does Notte Browser MCP work with Claude Desktop?
Yes. The Notte Browser MCP server follows the Anthropic MCP spec and is meant for MCP-native desktop clients like Claude Desktop and Windsurf. Setup and current status live in docs.notte.cc/integrations/mcp. For coding agents that run in a terminal, prefer the CLI + Skill path.
Do all three surfaces share the same cloud browser?
Yes. CLI, Skill, and MCP all authenticate with the same NOTTE_API_KEY, run against the same cloud browser infrastructure, and resolve to the same session profiles, vaults, and Functions. Mix and match across machines and clients - state follows your account, not the surface.
How is Notte different from Playwright MCP or @browsermcp/mcp?
Playwright MCP and @browsermcp/mcp run a local browser on your machine. They are free and fine for simple, public, low-volume tasks. They do not give your agent stealth fingerprints, residential proxies, CAPTCHA solving, or persistent cloud session profiles. Notte runs the browser in the cloud with all of that built in, which is what you need the moment a site detects bots or you want to resume an authenticated session across runs.
Is it open source?
The Notte CLI (github.com/nottelabs/notte-cli, MIT) and the Notte Skills bundle (github.com/nottelabs/notte-skills, MIT) are open source. The Notte SDK and MCP server live in github.com/nottelabs/notte. The cloud backend that runs the browsers is a hosted service with a free tier and usage-based pricing.
Free tier, no credit card. Start with the CLI and the Agent Skill; add the MCP server if and when your client expects it.