Google quietly shut down Project Mariner on May 4, 2026. No public announcement. No blog post. Just removed.
The timing is hard to ignore. Google I/O 2026 kicks off May 19, where Google will unveil Remy, their new 24/7 personal agent. Project Mariner had to be cleared off the stage first.
What Mariner Actually Was
Launched in December 2024 and demoed at I/O 2025, Project Mariner was Google's attempt at a browser-based AI agent that worked by taking screenshots of webpages and visually "reading" them. Every click required a vision pass. Every form field needed OCR-style reasoning. The idea sounded clever on paper.
In practice, it was too slow, too expensive, and too error-prone at scale. Seventeen months of development, and the approach never got past its core limitation: screenshots are a terrible interface for automated systems.
Where the Technology Went
Google folded Mariner's tech into two products. Gemini Agent, their new task-automation layer, got some of the underlying models. Chrome also picked up an Auto Browse feature, though it's restricted to AI Pro and Ultra subscribers in the US only.
The first warning signs appeared in March 2026 when Google moved staffers off the Mariner team. By early May, the project was gone.
Why API-First Agents Won
The screenshot approach tried to make AI interact with the web the way humans do. But that was always the wrong abstraction. APIs exist for a reason: they're faster, cheaper, and more reliable than parsing pixels.
OpenClaw, with over 368K GitHub stars, takes the opposite approach. It connects to services through APIs and the Model Context Protocol (MCP), not by staring at rendered HTML. When you need your agent to interact with a tool, you give it a direct connection, not a magnifying glass.
That difference matters for reliability. A screenshot-based agent fails when a button moves 20 pixels. An API-based agent keeps working until the API changes, which is versioned and documented. If you're running an OpenClaw instance through ClawHosters, your agent talks to external services directly. No pixel-parsing middleman.
Google seems to agree now. Remy, their replacement, is expected to lean heavily on structured tool calls rather than visual browsing. The screenshot era lasted 17 months. The API-first approach keeps winning.