Mouse-1

Mouse-1 is a decision model from Mouse, offered through an API, that sits inside an agent harness and handles the many small decisions an agent makes during a task. It takes the agent's state and a set of options and returns a probability for each, so the main frontier model can focus on reasoning and code.

What It Is

Mouse-1 is a decision layer for agent harnesses. Instead of generating text, it scores options supplied by the caller, such as routing, gating or verification choices. The harness uses the returned probabilities to decide what to do next, while a larger model plans and writes code and tools execute actions.

How It Works

A single request can answer up to 128 questions, and Mouse says responses take about 100 ms. The site shows example classifications on images and screens: content moderation, document routing and browser-state detection (for example error, loading, ready or login required). Mouse states that requests are scored and dropped with zero data retention, and that nothing is used for training.

Vendor Benchmarks

Mouse reports 89.1% average decision accuracy on unfamiliar tools and repositories, and a 0% decision flip rate when instructions are planted in the input, compared against models such as d3, Clef, OpenAI Decisions and Jev. Because Mouse-1 only picks from the options it is given, the vendor says it is hard to inject. These figures come from Mouse's own tests.

Related Mouse Products

The wider Mouse site also describes a mobile app for supervising coding agents, with plan review, notifications when an agent needs a decision, and cloud agent workspaces running in Firecracker microVMs.