Accessibility technology / computer-use agents
Fluent
An accessibility-first AI agent designed to operate a computer through inputs such as voice, gaze, breath, sign, or a single tap.
A computer-use agent should adapt to the person, not demand one approved input device.
The problem
Desktop interfaces are still organized around precise keyboard and mouse actions. That can create unnecessary friction or exclusion for people with repetitive strain injury, motor impairments, temporary injuries, or situations where conventional input is impractical.
The approach
Fluent starts from intent rather than a fixed sequence of clicks. The product interprets an accessible input, reads the state of the interface, plans an action, and exposes enough feedback for the user to confirm, correct, or stop what happens next.
Public capabilities
- Publicly described input modes include voice, gaze, breath, sign, and a single tap.
- A desktop sandbox demonstrates browser, writing, scheduling, and summarization tasks.
- The public product offers a Windows download and a live-demo path.
- The product is being developed accessibility-first rather than retrofitting accessibility after launch.
Boundaries
- Input support, task reliability, and compatibility should be evaluated per release and per user setup.
- Task completion alone is not sufficient evidence of accessibility; correction burden, recovery, consent, and user control also matter.
- Public user or deployment counts remain self-reported unless an independent source audits them.
Straight answers
What is Fluent?
Fluent is an accessibility-first computer-use agent that translates inputs such as voice, gaze, breath, sign, or a single tap into desktop actions.
Who is Fluent designed for?
It is designed for people who cannot always rely on a keyboard and mouse, including people with repetitive strain injury or motor impairments, while remaining useful to a wider audience.
How should an accessibility agent be evaluated?
Evaluation should measure successful task completion together with input effort, correction burden, recovery, safety, latency, and the user's ability to understand and stop actions.