Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Everyday tasks

Luma: an NVDA add-on that describes the screen

Written November 12, 2026

NVDA tells you what a control is. Luma tells you what the screen looks like: it sends a capture to an AI model you choose and reads back a description, and in Live Mode it speaks only when something on the screen actually changes.

What Luma adds

Luma is an add-on for the NVDA screen reader that brings AI-powered visual description, screen analysis and conversational chat to blind users. It can capture the entire screen, the current application window, or a specific navigator object and send the capture to a vision-capable AI model for description. It also processes clipboard text and images, frames from a webcam, pages from a scanner, and the file selected in File Explorer, including images, PDFs and PowerPoint slides. It works with NVDA 2025.1 through 2026.1 and is distributed under the GNU General Public License v2 or later.

Live Mode: narrate the change, not the screen

The feature that separates Luma from a describe-this button is Live Mode: continuous change-only narration of the screen. Instead of re-describing everything on a timer, it speaks when the screen changes. A second live form lets you ask a question, such as what the slide title is, and get an answer whenever the answer changes. For a presentation, a download, or a slowly updating dashboard, that is the difference between a running commentary you can follow and a wall of repeated description.

What it sends where

This is the part to set up with your eyes open. A description means a screenshot leaves the machine unless you choose otherwise. Luma's built-in providers are OpenAI, Anthropic (Claude), Google Gemini and OpenRouter, and any custom OpenAI-compatible endpoint can be added; a capture sent to one of those services is processed under that service's terms. Luma also lists Ollama as a built-in provider and bundles an on-device inference engine, so a fully local route exists for screens you do not want to send out. The choice is per provider and per skill: reusable skills are defined as simple Markdown files with frontmatter, and each skill can name its own provider, model and temperature. Ten shortcut keys can each be mapped to a skill and scope. The everyday keys are NVDA+Shift+G to describe the screen, NVDA+Shift+R to describe the navigator object, and NVDA+Shift+Space to cycle skills.

One feature deserves a plain warning label. Interaction Mode lets the AI operate an inaccessible interface for you: it takes screenshots, clicks, types and presses keys until the task is done. That is real control handed to a model, working natively with Anthropic and Gemini and, through a prompted protocol, with OpenAI-compatible vision models. It is powerful exactly where an interface gives a screen-reader user no other route, and it should be switched on task by task, not left running.

Getting it

Download the latest luma-<version>.nvda-addon file from the repository's Releases page, open it, and NVDA will prompt to install; then restart NVDA when prompted. Set up a provider under NVDA menu, Tools, Luma, Providers, choose defaults under Tools, Luma, Settings, and the add-on is ready. The project is young and small: as read for this page, the repository shows a handful of commits, four releases and a single star. That is not a reason to dismiss it, and it is a reason to keep your NVDA setup backed up and to test Luma on screens that do not matter first.

Limits

Related playbooks

Sources

Back to: Everyday playbooks hub · GeoVisA11y: maps a screen reader can explore · CodeA11y: a coding assistant that watches accessibility