Project Mariner

Project Mariner is Google DeepMind's experimental browser agent, built on Gemini 2.0, that takes over your Chrome browser to complete web tasks on your behalf. Announced in December 2024 as a research prototype, Mariner reads pixels, web elements, text, and forms simultaneously, then acts on what it sees. It scored 83.5% on the WebVoyager benchmark at launch, beating the competition by a meaningful margin. By 2026, the core browser-automation capability has shipped under the name Chrome Auto Browse inside Google AI Pro and Ultra plans. This review covers what Mariner can do, where it falls short, how it stacks up against OpenAI Operator and Anthropic Computer Use, and whether Google's version of the autonomous browser is worth your subscription dollars.

When Google showed off Project Mariner in December 2024, it looked like a direct shot at OpenAI’s Operator. A Chrome extension powered by Gemini 2.0 that reads your browser screen, reasons about what it sees, and then clicks, scrolls, and types to get things done. The demo was slick. The benchmark number, 83.5% on WebVoyager, was the highest any single agent had posted at the time. For a brief stretch it felt like Google had shown up to the autonomous-agent race with something real. This review looks at what Project Mariner actually delivered, how it evolved from research preview to shipping product, and whether it’s worth paying for in 2026.

Quick verdict

Mariner started stronger than almost anyone expected and Google followed through by shipping the underlying capability inside Google AI Pro and Ultra rather than letting it rot in the research lab. It’s not fast, it doesn’t support multiple tabs at once, and there’s no developer API. But if you pay for Google’s AI subscription and want a browser agent that understands complex pages without falling apart, Mariner is the most mature option Google has ever shipped in this space.

What is Project Mariner, exactly?

Google DeepMind announced Project Mariner on December 11, 2024, as part of the Gemini 2.0 launch wave. The framing was explicit: this is Google’s answer to the question of what happens when a capable multimodal model gets direct access to your browser.

The mechanism is a Chrome extension. Install it, describe a task in plain English, and Mariner reads everything on your active tab, including the raw pixels, the underlying HTML elements, any forms, images, or embedded content, and then builds a plan. It doesn’t simulate keystrokes the way older RPA tools do; it reasons about what the page contains and decides which elements to interact with. That distinction matters because it means Mariner can handle pages that would break a traditional script the moment a designer changed a button label or moved a field.

At the time of launch, Google was careful to frame this as a research prototype. Access went to trusted testers only. The stated caveats were honest: the agent was “not always accurate and slow to complete tasks today.” Those caveats haven’t fully gone away by 2026, but the product has moved past the prototype stage. The browser-automation capability is now bundled into Google AI Pro and Ultra under the name Chrome Auto Browse, which means paying subscribers can use it without applying for a waitlist.

The Gemini 2.0 foundation is what separates Mariner from earlier attempts at browser automation. Gemini 2.0 was built from the start to reason across modalities simultaneously. It doesn’t convert a screenshot to text and then reason about the text; it processes the visual, structural, and textual information at the same time. That’s why the WebVoyager benchmark number came in as high as it did, and it’s why Mariner handles visually complex pages better than agents that rely on extracted text alone.

The features that drove the December 2024 launch

Chrome extension that takes over your browser

The delivery mechanism is deliberately simple: a Chrome extension that connects your browser session directly to the Gemini 2.0 model. When you give Mariner a task, it gains control of the active tab. It can click any element, fill any form, scroll the page, and navigate to URLs. It does not open new tabs on its own, and it cannot see or interact with tabs you have open in the background.

That single-tab constraint is both a safety feature and a real limitation. It’s a safety feature because it keeps the agent’s blast radius contained: if something goes wrong, it’s contained to the one page you pointed it at. It’s a limitation because real-world workflows frequently involve more than one window. Comparing prices across three travel sites, for example, requires Mariner to do them sequentially rather than in parallel, which is slower than just doing it yourself.

Gemini multimodal as the brain

Mariner doesn’t work by extracting the DOM and feeding it to a language model as text. It processes the browser screen the way a person does, reading layout, color, position, and content all at once. Gemini 2.0 was Google’s first model built with native multimodal input at scale, and Mariner was purpose-built to use that capability.

What this means in practice: Mariner can read a data table rendered as an image, understand a visual form where fields aren’t explicitly labeled in HTML, and handle pages where the interactive elements are built in JavaScript and don’t show up in a traditional DOM traversal. That’s a genuine advantage over agents that rely purely on accessibility trees or HTML extraction.

The 83.5% WebVoyager benchmark score at launch put Mariner ahead of every other publicly tested browser agent at the time. WebVoyager tests real-world web tasks across diverse sites; it’s not a narrow synthetic benchmark. Scoring above 80% as a single-agent system was notable.

Sandboxed task execution

Google made a deliberate architectural decision to limit Mariner to the active tab. The agent can’t access your local filesystem, can’t open background tabs without your permission, and can’t interact with browser extensions other than itself. This isn’t a technical limitation; it’s a policy boundary Google chose.

The effect is that Mariner behaves more like an attentive assistant working on the one page you pointed at than a general-purpose computer-control agent. You’re not handing over your entire machine. You’re handing over one browser tab for the duration of a specific task.

Take-over and supervision

The most important safety feature in Mariner is the confirmation gate. Before Mariner completes any action that Google classifies as sensitive, specifically purchases, form submissions that send data to a server, and account changes, it pauses and asks you to confirm. You can see exactly what it’s about to do before it does it.

This “human in the loop” model is different from fully autonomous agents that run to completion without interruption. Mariner is designed to be supervised. That’s appropriate given where browser agents are in 2026: capable enough to do useful work, but not reliable enough to trust blindly with your credit card and login sessions. The tradeoff is that you can’t set Mariner loose and walk away. You need to be available to approve the sensitive steps.

Integration with the Gemini ecosystem

Because Mariner runs inside Chrome and connects to Google’s AI infrastructure, it has natural advantages when the task involves Google products. Summarizing an email thread and creating a calendar invite from it, pulling data from Google Sheets into a web form, cross-referencing Google Search results: these workflows run more smoothly than equivalent tasks on third-party sites, because Mariner’s authentication and context are already present in your browser session.

This isn’t a unique capability, but it’s a real one. If your work already lives in Google Workspace, Mariner has less friction than a cloud-based agent that needs to authenticate into your accounts from scratch.

Pricing

Project Mariner started with no pricing because it had no public access. The December 2024 launch put it in the hands of trusted testers at no charge, but also with no guarantee of continued availability.

By 2025, Google folded the browser-automation capability into its paid subscription tiers. As of May 2026, Chrome Auto Browse, which is Mariner’s core browser-agent feature shipped as a product, is available to Google AI Pro and Google AI Ultra subscribers.

Google AI Pro runs $19.99 per month and includes 5 TB of Google One storage alongside the full suite of Gemini AI features. Google AI Ultra sits at a higher tier with 30 TB of storage and higher rate limits across all AI features including Chrome Auto Browse.

There is no free tier for browser automation. The Gemini app itself has a free tier, but agentic browser control is locked to paid plans. That’s consistent with how OpenAI handles Operator (ChatGPT Plus/Pro required) and how Anthropic positions Computer Use (API access with per-token billing).

If you’re already paying for Google One at the Pro level for storage, you’re getting Mariner’s browser agent essentially bundled in. If you’re evaluating this purely as a browser-agent tool without caring about the storage, $19.99 per month is on the lower end of the market compared to Operator’s pricing, which makes it competitive on price even if the feature set isn’t identical.

There’s no developer API for Mariner’s browser-control primitives. If you want to build autonomous browser workflows programmatically, you’ll need to look at Anthropic Computer Use via the API or a framework like Browser Use that doesn’t require a subscription.

Where Mariner wins and where it doesn’t

Mariner is genuinely good at tasks that require reading complex web pages and extracting or filling in information. Hotel comparison across a single booking site, reading a product spec table and answering a question about it, filling out a multi-field contact form: these work well and feel close to having a capable assistant do the mechanical clicking for you.

Where it struggles is speed. Multi-step tasks take longer than you’d expect. Mariner reasons before each action, which means a five-step workflow can take two or three minutes in wall-clock time. If you’re doing something that takes you thirty seconds manually, Mariner probably isn’t saving you time; it’s just saving you attention.

The single-tab constraint is a real bottleneck for comparative research tasks. “Find the cheapest flight across Google Flights, Kayak, and Expedia” sounds like a perfect agent task, but Mariner has to visit each site sequentially. By the time it finishes the third site, prices may have changed.

Bot-detection is an ongoing problem. Sites like Ticketmaster, some airlines, and checkout pages with aggressive CAPTCHA systems will stop Mariner just as they stop traditional automation scripts. Google doesn’t publish a list of sites where Mariner reliably works versus fails.

Who Mariner is built for

Mariner fits best with people who are already inside the Google ecosystem and find themselves doing repetitive browser tasks throughout their workday. Think: someone who regularly pulls information from web sources into Google Docs, or fills out vendor forms with data that already exists in a spreadsheet, or needs to compare options across a handful of structured websites.

It’s less suited to power users who want to build custom automation pipelines. There’s no API, no webhook support, no way to chain Mariner tasks into a larger automated workflow without manual intervention. For that use case, Anthropic Computer Use with a custom agent loop or a framework built on browser-control primitives gives you more control.

It’s also not ideal for anyone who needs the agent to run unattended. The confirmation-gate model means you need to be present. If you want “set it and forget it” automation for low-stakes tasks, something like Manus with its asynchronous task model is a better fit.

Mariner vs the alternatives

OpenAI Operator is the most direct comparison. Both launched within weeks of each other, both require a paid subscription, and both use a browser as the interface between the AI model and the web. The key architectural difference is that Operator runs a cloud-hosted browser session, meaning it doesn’t take over your screen and doesn’t have access to your existing login sessions. Mariner runs directly in your Chrome instance, which gives it access to cookies and sessions you already have but also means it visually occupies your browser while working. Operator’s cloud model is better for unattended tasks; Mariner’s local model is better for tasks that require your existing authentication. On raw benchmark performance at launch, Mariner scored higher. In practice, both agents have similar failure modes: slow execution, bot-detection walls, and difficulty with truly dynamic JavaScript-heavy pages.

Anthropic Computer Use takes a different approach entirely. Instead of a purpose-built browser agent, it’s an API capability that lets Claude see and control an entire desktop environment, including the browser, file system, and applications. It’s far more flexible than Mariner: you can build custom agent loops, run it headlessly in Docker, and use it for non-browser tasks. The tradeoff is that it’s a developer tool, not a consumer product. There’s no polished extension to install; you’re building the scaffolding yourself. If you want browser automation out of the box with a UI, Mariner wins on convenience. If you want to build something custom or automate tasks that go beyond the browser, Computer Use is the more powerful foundation.

Manus is an async-first agent that handles longer-horizon tasks by running in the background and reporting back when done. It can browse the web as part of multi-step research or task-completion workflows, but it’s not exclusively a browser agent. Manus handles the kind of tasks where you want to describe a goal, come back in an hour, and find the work done. Mariner isn’t built for that pattern. If the task you’re automating is a sequence of web interactions that needs to complete in a single session while you watch, Mariner is more appropriate. If it’s a longer research or analysis task you want to run overnight, Manus is a better choice.

For users who want to know more about what agents can do beyond the browser, the best AI agents for coding roundup covers the agentic landscape for development workflows specifically.

Getting started

To use Chrome Auto Browse (Mariner’s shipped form), you need a Google AI Pro or Ultra subscription. Sign up at one.google.com, select AI Pro at $19.99 per month, and then access the feature through the Gemini app in Chrome.

The extension installs from the Chrome Web Store once your subscription is active. After installation, open the Gemini sidebar in Chrome, type a task description in plain language, and click to start. Mariner shows its reasoning and planned actions before executing each step, so you can intervene early if it’s heading the wrong direction.

Start with low-stakes tasks to get a feel for how it handles your most-visited sites. A good first task is something like “go to [site] and find the return policy” or “fill out this contact form with my standard info.” Avoid testing it on checkout flows or account settings until you’ve confirmed it handles the simpler tasks correctly.

Google’s own documentation recommends staying present and keeping tasks scoped. The more specific you are in the initial prompt, the better Mariner performs.

The bottom line

Project Mariner had a genuinely strong debut in December 2024, with a benchmark score that outperformed the competition and a safety model that was more thoughtful than most. What matters more in 2026 is that Google followed through: the capability shipped inside a real product rather than disappearing into the research lab backlog.

It’s not perfect. It’s slow. It doesn’t do multiple tabs. It has no developer API. But for a Google subscriber who wants browser automation without writing a line of code, Mariner is the most accessible entry point in the market. If you’re already on Google AI Pro, you should try it. If you’re evaluating it as a standalone purchase, compare your expected use against what OpenAI Operator offers at its price point before committing.

Key features

• Chrome extension that takes over the active browser tab to complete multi-step tasks
• Gemini 2.0 multimodal brain reads pixels, web elements, text, forms, and images simultaneously
• Sandboxed execution: agent is limited to the currently active tab and cannot access other tabs or local files
• Human-in-the-loop confirmation gates for sensitive actions such as purchases or form submissions
• 83.5% score on the WebVoyager benchmark for end-to-end web task completion
• Integration with the broader Gemini ecosystem and Google One subscription tiers
• Real-time task status display so users can monitor and interrupt agent actions

Pros and cons

Who is Project Mariner for?

• Price comparison across hotel booking sites without manual tab-switching
• Filling out multi-step web forms using data the user provides once
• Researching and summarizing information across several websites in a single task
• Automating repetitive Google Workspace actions like creating calendar invites from email threads

Alternatives to Project Mariner

If Project Mariner isn't quite the right fit, the closest alternatives are openai-operator , anthropic-computer-use , and manus . See our full Project Mariner alternatives page for side-by-side comparisons.