--- title: "Private Anti-Detect Browser for AI Agents — Donut Browser" description: "Run Claude Code, Cursor, Codex or any MCP agent in isolated profiles, each with its own fingerprint, proxy and cookies. No AI model inside, zero telemetry." url: "https://donutbrowser.com/use-cases/ai-agents/" --- 1. [Home](https://donutbrowser.com/) 2. [Use Cases](https://donutbrowser.com/use-cases) 3. The private browser for AI agents For Claude Code, Cursor, Codex and any MCP client # The private browser for AI agents Each agent task runs in its own profile, with its own fingerprint, proxy and cookies. Donut has no AI model of its own, and the app sends no telemetry. You bring the agent. The profiles stay on your disk. Download Donut Browser Read the MCP docs - No AI model inside - Zero app telemetry - macOS, Windows and Linux Compare Donut profiles One identity An agent gets one task and opens 6 Donut profiles, one per seller account, each on its own site. Every profile has its own operating system, fingerprint, time zone, exit IP and cookies, so a site check sees 6 different people. In the comparison, all tabs share one fingerprint and one IP, and the same check links all 6 accounts. Nordlicht Shop: macOS 15, Europe/Berlin, exit IP 185.72.14.9, own cookies. Lyon Épicerie: Windows 11, Europe/Paris, exit IP 91.168.20.44, own cookies. Osaka Craft: Windows 10, Asia/Tokyo, exit IP 133.106.8.12, own cookies. Lakeside Goods: macOS 14, Europe/Madrid, exit IP 88.12.140.61, own cookies. Bottega Roma: Windows 11, Europe/Rome, exit IP 93.41.77.150, own cookies. Wisła Home: Windows 11, Europe/Warsaw, exit IP 83.24.61.201, own cookies. ## Bring the agent you already use Any AI browser agent that speaks MCP can drive Donut. Scripts can drive the same profiles with Playwright or Puppeteer. The app writes the MCP config for these 22 clients in one click - Claude Code - Claude Desktop - Cursor - Codex - VS Code - Zed - Windsurf - Gemini CLI - GitHub Copilot CLI - Cline - Cline CLI - Goose - OpenCode - Kilo Code - Kiro CLI - Kimi Code - Grok Build - Antigravity - Mastra Code - MCPorter - Pi - fx ### Remote MCP Donut’s browser MCP server is one URL and a dmk\_ credential. It has more than 100 tools for profiles, proxies, fingerprints, cookies and the browser itself. The desktop app opens an outgoing connection, so no port opens on your computer. JSON Copy ``` { "mcpServers": { "donut-browser-remote": { "url": "https://api.donutbrowser.com/api/mcp", "headers": { "Authorization": "Bearer dmk_…" } } } } ``` Solo and up. Browser tools need Pro. ### Playwright or Puppeteer over CDP Start a profile from the local REST API and connect to its debug port on 127.0.0.1. The wayfern npm package adds humanized mouse, typing and scroll to Playwright. Node.js Copy ``` import { chromium } from "wayfern"; const res = await fetch( `http://127.0.0.1:10108/v1/profiles/${id}/run`, { method: "POST", headers: { Authorization: `Bearer ${token}` } }, ); const { remote_debugging_port } = await res.json(); const browser = await chromium.connectOverCDP( `http://127.0.0.1:${remote_debugging_port}`, ); ``` Profile management is free. Starting profiles from code needs Pro. ## What your agent’s browser sends, and to whom Each kind of data, where it goes, and which parts pass through Donut’s servers. Hosted AI browser: from Aside’s privacy policy and help pages, retrieved 14 Oct 2026. Aside comparison What leaves your computer, in two diagrams. Donut Browser: your agent (Claude Code, Cursor, Codex) drives Donut through the local API and CDP on 127.0.0.1 and uses the model you choose. Profiles, cookies, history and fingerprints stay on your disk. Websites get the pages your agent visits, through each profile’s proxy. Donut sends nothing to model providers and has no telemetry. Optional sync is end-to-end encrypted, and remote MCP commands pass through Donut’s relay. Hosted AI browser, based on Aside’s privacy policy: with hosted models, model providers get prompts, page snapshots, screenshots and files; an analytics vendor gets your email, plan and device model, on by default; encrypted history and cookies go to the vendor’s cloud when sync is on. | Data | Goes to | Details | | --- | --- | --- | | Stays on your computer | | App usage and crash data | Nobody | The desktop app has no telemetry, analytics or crash reporting. | | Profiles, cookies and history | Your disk | No account is needed to use the app. Cloud sync is off by default. | | Local API and CDP | This computer | The REST API and each debug port listen on 127.0.0.1 only. | | Leaves your computer | | Prompts and page text | Your agent’s model provider | Donut has no model and sends nothing to model providers. Your agent sends what it reads to the provider you chose. | | Remote MCP calls | Donut’s relay, then your desktop app | Calls and results pass through api.donutbrowser.com over TLS. They are not end-to-end encrypted. Proxy passwords are redacted, and tools that read local files or export secrets are refused. | | Synced profiles | Donut’s sync server, or your own | Sync is optional. With end-to-end encryption on, the key comes from your password (Argon2id, AES-256-GCM). The server is self-hostable. | What AI browsers send to model providers ## Run one AI agent on multiple accounts Sites link accounts by device and network. Give each account its own profile, and the accounts your agent works in share no fingerprint, IP address or cookies. | Profile | Fingerprint | Exit IP | Timezone | Language | | --- | --- | --- | --- | --- | | Sales, rep 1 | macOS 15 · Apple M2 a41f 9c02 | 203.0.113.24 DE | Europe/Berlin | de-DE | | Sales, rep 2 | Windows 11 · RTX 4060 9c07 e1b8 | 198.51.100.7 US | America/Chicago | en-US | | Recruiting | macOS 14 · Apple M1 52be 07d3 | 192.0.2.81 GB | Europe/London | en-GB | | Vendor portal | Windows 10 · Iris Xe e3d9 6a15 | 203.0.113.140 FR | Europe/Paris | fr-FR | Illustrative data. Four accounts, four profiles. ### Fingerprint in the engine Wayfern, Donut’s Chromium-based engine, sets each profile’s fingerprint in C++, under the page’s JavaScript. Fingerprint editing and other operating systems’ fingerprints are on Pro. ### A proxy or VPN per profile HTTP, HTTPS, SOCKS4, SOCKS5 or WireGuard. DNS goes through the same route, and WebRTC leak protection keeps the real IP out of the page. ### Timezone and language from the IP Both follow the proxy’s exit IP. If they do not match, Donut does not start the profile. ### Cookies of its own Each profile keeps its own cookies and storage. Import or export them as JSON, Netscape or Playwright storage state. How to run an AI agent on multiple accounts ## AI agents for sales, recruiting, operations and research One real task per team, and why it runs in a private profile of its own. ### Sales The task: Before each call, the agent reads the prospect’s website and your CRM record, then writes a one-page brief. It stops before anything is sent. Why its own profile: Each rep’s CRM and email stay signed in to that rep’s profile, on that rep’s IP address. No agent works in another rep’s session. AI agents for sales ### Recruiting The task: The agent reads new applicants in your ATS, opens each public portfolio and drafts a shortlist for you to check. Why its own profile: Candidate data stays in a profile on your disk. Donut stores none of it. Only your agent’s model provider sees what the agent reads. ### Operations The task: Once a month, the agent signs in to each vendor portal, downloads the invoices and lists the renewals due in 30 days. Why its own profile: One profile per portal keeps the same login, device and IP every month, so each portal sees a returning user. ### Research The task: The agent runs the same search from three countries, saves every source and builds an evidence table with links. Why its own profile: Each profile has its own exit IP, timezone and language, so each search runs as a local reader’s would. ### Developers The task: Your code starts profiles from the local API and connects Playwright over CDP. Your agent reaches the same profiles over remote MCP. Why its own profile: The debug port listens on 127.0.0.1, so the local path stays on your computer. Automation with Playwright ## You stay in control of every run Agents ask for help in the app. Tasks you repeat become recipes that replay without a model. Solo and up ### Answer, take over or pause When an agent needs a choice, a login or a captcha, it calls ask\_human or request\_human\_help. The question appears on the Agent page in Donut. You answer it, take over the browser, or pause the agent. The console is kept in memory only. Claude Code works in its own Donut profile You take over the profile. Claude Code waits. Claude Code calls ask\_human. You answer on the Agent page. Claude Code continues and finishes the task Donut’s Agent page while Claude Code works in the Nordlicht Shop profile on seller.example. At the sign-in, you take over the profile and enter the one-time code. Claude Code waits and continues when you hand the profile back. It finds two March invoices and calls ask\_human to ask whether to download both. You answer on the Agent page, and it saves both. Pause agents stops every agent until you resume. Solo and up ### Record once, replay with no tokens Record a task from your own clicks and typing. Donut replays it step by step without a model, so a replay costs no tokens. Passwords are never recorded. Recipes are stored in Donut’s cloud. Solo saves recipes and schedules them on remote browsers. Running a recipe on demand, from an agent or on your desktop needs Pro. You do the task once. Donut records each step. The replay runs the steps with no model. 0 model tokens. A schedule can run it on remote browsers. You download last month’s invoices once in the Nordlicht Shop profile on seller.example, and Donut records each click as a recipe step. Passwords are never recorded. The recipe is saved in Donut’s cloud. A replay runs the five steps with no model, so the model token count stays at 0. Running a recipe now, from an agent or on your computer needs Pro. A schedule runs it on remote browsers from Solo. ## What each plan lets your agent do The app and its profiles are free. Agents connect from Solo, and drive the browser from Pro. ### Free $0 - The app and unlimited local profiles - Fingerprint, proxy and VPN per profile - Local API to manage profiles - No MCP ### Solo $29 / month - Remote MCP for profiles, groups, proxies, VPNs and extensions - Human-in-the-loop tools - Save recipes and schedule them on remote browsers ### Pro $99 / month - Browser control: launch, navigate, click, type, scroll, screenshot, evaluate - Run recipes on demand, from an agent or on your desktop - Start profiles from code over CDP - 200 remote browser hours a month Team adds shared profiles with locks and roles: $199 a month for 2 seats, then $99 a seat. Compare plans ## Questions, answered. Which AI agents work with Donut Browser? Any MCP client. The app writes the MCP config in one click for 22 clients, including Claude Code, Claude Desktop, Cursor, Codex, VS Code, Zed, Windsurf and Gemini CLI. Scripts can drive the same profiles with Playwright or Puppeteer over CDP. Setup steps are in the [Remote MCP docs](https://donutbrowser.com/docs/remote-mcp). Does Donut Browser include an AI model? No. Donut has no model and sends nothing to model providers. You bring your own agent, and it talks to the model provider you chose. What does Donut see of my agent’s data? On the local path, nothing: the REST API and CDP traffic stay on 127.0.0.1. With remote MCP, each tool call and its result pass through Donut’s relay at api.donutbrowser.com on the way to your desktop. That traffic uses TLS but is not end-to-end encrypted, and proxy passwords are redacted. Recipes are stored in Donut’s cloud. The app itself sends no telemetry. Can websites detect my AI agent? Donut controls what a site reads about the device and the network: each profile has its own fingerprint, IP address, timezone and language. Sites can still judge behavior. A 2026 study ( [FP-Agent](https://arxiv.org/html/2605.01247v1)) detected all seven AI agents it tested by how they acted. Keep agents at a human pace, use the wayfern package for humanized input, and follow each site’s rules. [Why websites flag AI agents as bots](https://donutbrowser.com/ai-agent-bot-detection) explains the signals. Can one AI agent work in many accounts? Yes. Give each account its own profile. The agent opens the profiles it needs and works in each one, and the accounts share no fingerprint, IP address or cookies. Automation is rate limited on every plan. Is Donut Browser free for AI agents? The app, unlimited local profiles and the local API for managing profiles are free. Remote MCP starts with Solo at $29 a month. Solo also saves recipes and schedules them on remote browsers. Browser control over MCP, running recipes on demand or from an agent, and starting profiles from code need Pro at $99 a month. How is Donut different from an AI or agentic browser? An AI browser ships with its own assistant, and every task runs with the browser’s one device fingerprint and IP address. Donut has no assistant. The agent you already use drives Donut profiles, and each profile has its own fingerprint, proxy and cookies. The app needs no account and sends no telemetry. How does my agent ask me for help? It calls a human-in-the-loop tool such as ask\_human or request\_human\_help. The question appears on the Agent page in Donut, where you answer it, take over the browser or pause the agent. These tools are included from Solo. Can agents run on remote browsers? Yes. Pro includes 200 hours a month of remote browsers on Windows, macOS and Linux machines in Donut’s cloud. Scheduled recipes run there from Solo. A remote profile needs sync without end-to-end encryption. Does Donut Browser work on Linux? Yes. Donut runs on macOS, Windows and Linux, with .deb, .rpm and AppImage packages for Linux. MCP clients and Playwright scripts on Linux connect the same way. ## Give your agent a private browser Download the free app and make a profile for each account. Connect your agent with Solo, and let it drive the browser with Pro. Download Donut Browser Compare plans More use cases Web Scraping & Automation Privacy Protection Multi-Account Management