Skip to content

project mariner is gone

github-actions[bot] edited this page Sep 3, 2026 · 1 revision

Project Mariner is gone: what replaced it

Project Mariner is gone. Google's own landing page for the experiment records the service as shut down on 4 May 2026, and the capability did not die with it: it was folded into Gemini Agent inside the Gemini app, into Chrome's auto browse feature, and into a computer use tool in the Gemini API that developers can call today. If you subscribed to Google AI Ultra partly for Mariner, those three are where your money now points. If you wanted the capability without a Google subscription attached to it, the last section of this page is the part written for you.

A disclosure before anything else: this page lives on the wiki of AIHawk, an open-source agent in the same category, which appears in that last section with the conflict stated. Every date and claim here was checked against Google's pages or press coverage of them on 2026-09-03.

What Mariner was

Google announced Project Mariner in December 2024 as one of the first AI systems designed to control a web browser on a user's behalf. It worked the way most of the category still works: take a screenshot, identify the text and the buttons, click and type the way a person would, repeat until the task is done. Access came through the Google AI Ultra subscription at $250 per month, which made it one of the most expensive ways to try a browser agent that has ever been offered.

It was always labeled an experiment, and Google was explicit that it was a research vehicle rather than a product. That label turned out to be a precise description of its life expectancy.

How it ended

There was no blog post and no deprecation window. Press coverage in early May 2026 noticed that Mariner's landing page had quietly changed to say the service ended on 4 May 2026, and that the page now points users at other Google products for the same work: Gemini Agent for complex tasks, and AI Mode in Search for research-shaped ones. The shutdown landed about two weeks before Google I/O 2026, which is consistent with a consolidation ahead of the event rather than a retreat from the category, because the capability itself shipped again in three places.

If this story sounds familiar, it is because OpenAI ran the same arc first: Operator shut down in August 2025, its successor agent mode was removed a year later, and the Atlas browser closed the same month. The pattern is documented on Is OpenAI Operator still available?, and the lesson transfers unchanged: a hosted agent is a feature inside someone else's product strategy, and it moves when the strategy moves.

Where the capability went

Three heirs, each a different shape, all verified against Google's material this session.

Gemini Agent, for subscribers

Gemini Agent arrived in the Gemini app on 18 November 2025, alongside Gemini 3 Pro, as an experimental feature for Google AI Ultra subscribers in the US. Google's own framing was direct about the lineage: it "brings what Google has learned with Project Mariner to the Gemini app". It handles multi-step tasks that mix live web browsing with Google's own tools - inbox triage, reminders, research tasks like pricing a rental car within a budget - and it asks for confirmation before sending emails, making purchases, and other consequential actions. During 2026 an Agent Mode toggle spread further down the subscription tiers.

This is the closest thing to "Mariner, continued" for a person rather than a developer, with one honest caveat: it is an agent that lives inside Google's assistant and acts across Google's surfaces, not a general watch-it-drive-any-site tool you aim wherever you like.

Auto browse in Chrome, for the browser itself

On 28 January 2026 Google began rolling out auto browse, a Gemini 3-powered agent inside Chrome, to AI Pro and AI Ultra subscribers in the US. Describe a chore - fill this form, compare these prices, book this slot - and Chrome opens a visually distinct tab, drives the page itself, and shows a step-by-step action log in a side panel. There is a daily limit on agentic actions, purchases and social posts require manual confirmation, and a Take over task button hands the tab back to you at any point. It reached Android later in the year.

Auto browse is the most literal answer to "what replaced Mariner", because it is the same capability in the same browser, now shipped as a Chrome feature instead of a standalone experiment.

The computer use tool, for developers

The Gemini API kept the machinery and dropped the product. Google introduced a dedicated Gemini 2.5 Computer Use model on 7 October 2025, built on Gemini 2.5 Pro, and described it at the time as primarily optimized for web browsers and not yet optimized for desktop OS-level control. Since then the offering has evolved from a dedicated model into a computer use tool that current mainline models call: the docs today recommend gemini-3.8-flash, list several other Gemini 3-family models, and keep gemini-2.5-computer-use-preview-10-2025 as the legacy entry.

The loop is the standard one for the category: you send a screenshot and a task, the model returns a suggested UI action with coordinates normalized to a 0-999 grid, your code executes it and sends back a fresh screenshot. Browser control is the primary target, with Android-optimized mobile support and OS-level cursor commands for desktop. Safety is configurable per category - financial transactions and account creation can be set to require confirmation - and prompt injection detection is available as an opt-in. Worth knowing before you build on it: Google's docs still label Computer Use a Preview capability that "may contain errors and security vulnerabilities", in the docs' own words.

If you want the capability without the subscription

Everything above requires either a Google subscription or a Google API bill, and all of it can be reorganized again, as 4 May demonstrated. The open-source column of this category runs on your machine with a model key you choose, and it survives vendor strategy changes by not having a vendor.

The maintained options, with stars read from each repository on 2026-09-03: browser-use (~112k stars, MIT) is the adoption leader and drives Chromium-family browsers; Skyvern (~23k, AGPL-3.0) takes a vision-first approach on Playwright; Agent S (~12k, Apache-2.0) operates the whole desktop, which makes it the nearest open relative of Mariner's screenshot generality; and AIHawk (~30k, MIT), our own project, pairs the agent with a Firefox patched at the source level so that what a page inspects looks like a normal desktop browser rather than an automation build. Ours is the one claim on this page we cannot make neutrally, so weigh it accordingly, and note its limits: Windows and Linux only, and you bring the model.

None of these, ours included, promises a site will not push back. That boundary belongs to every agent in the category and has its own page. The fuller landscape, hosted and open, is on OpenAI Operator alternatives and Open-source computer-use agents, and the two big-vendor APIs that survived their consumer products are compared on Gemini computer use vs Claude computer use.

Short answers to the questions that lead here

Is Project Mariner still available? No. Its landing page records the service as ended on 4 May 2026.

Did Google announce the shutdown? Not with a blog post. The landing page changed, press coverage noticed, and the page now redirects users to Gemini Agent and AI Mode.

What replaced Project Mariner? For subscribers, Gemini Agent in the Gemini app (since 18 November 2025) and auto browse in Chrome (since 28 January 2026). For developers, the computer use tool in the Gemini API.

Was Mariner folded into Gemini? Yes, in Google's own words: Gemini Agent "brings what Google has learned with Project Mariner to the Gemini app", and the Gemini API carries the browser-control tooling.

Do I still need Ultra to get the capability? Auto browse reaches AI Pro as well as Ultra in the US; Gemini Agent launched Ultra-first. The API route is pay-per-token, and the open-source route needs no Google subscription at all.

Is there an open-source equivalent? Several maintained projects do the same describe-a-task, watch-the-browser-work loop on your own machine; see the section above.

See also: Is OpenAI Operator still available? for the same arc at OpenAI, Gemini computer use vs Claude computer use for the developer tooling that outlived both consumer products, and Choosing an AI browser agent for how to pick a replacement that will not need replacing.

Sources


Maintained by the team behind AIHawk, an open-source AI web agent. We benefit when readers conclude that hosted agents are impermanent, which is exactly why every date above traces to Google's own pages or to coverage of them rather than to our say-so.

Alternatives and Comparisons

When the Agent Gets Blocked

Using the Agent

Clone this wiki locally