What is the toolkit: text version of the intro video
AI for Accessibility Collective. Video length 3 minutes 37 seconds. Text version written 3 November 2026.
About this text version
This is the whole video in text, for reading with a screen reader. It follows the video from start to finish, in five scenes. Each scene starts with its time in the video and a short description of what is on screen. Then come the words as they are heard, each line starting with who is speaking. "On screen" lines describe text and pictures that the spoken words do not cover.
One narrator speaks throughout. The video also uses real demo recordings of the toolkit, with the real speech heard during those demos. Each voice is identified where it speaks. The headset and phone views in scene 4 are simulated: working web pages, not real devices.
Scene 1 of 5. The vision and the three agents. From 0:00 to 0:54.
On screen: made-here slides on a warm cream background, with dark green text, a serif typeface for headlines and white cards with soft shadows.
On screen: a headline: "Many websites and apps still put barriers in front of disabled people."
Narrator: Many websites and apps still put barriers in front of disabled people. What's new now is agentic AI.
On screen: the headline changes to "What's new now is agentic AI." Two cards appear side by side. The first, headed "An agent uses a website", shows a simple web page with a button reading "Next step", and the words "Together with you: an AI agent can use a website with you." The second, headed "A coding agent builds an app", shows the request "Build me an app with large buttons", and the words "From plain words: a description becomes a working app."
Narrator: An AI agent can navigate a website together with you, and a coding agent can build an app from a description just using plain words.
On screen: "One day very soon, people may be able to build what they need for themselves."
Narrator: So, one day very soon, people may be able to build what they need for themselves. We started more simply, though, building code components that developers can pick up and use today.
On screen: "We started more simply: code components developers can pick up and use today."
On screen: a new headline, "Three agents work together with the person", and a row of four cards joined by arrows, appearing one at a time as each is named. The first card, dark green, reads "You: say what you need." Then "Ability profile agent: keeps a record of what you've asked for." Then "Adapter agent: changes a page or an app to match." Then "Verification agent: checks the result against your request." A line runs from the last card back to the first, marked "Back to you: the result comes back to you."
Narrator: Our toolkit has three agents that work together with the person. First, the Ability profile keeps a record of what you've asked for. Next, the Adapter agent changes a page or an app to match. And finally, the Verification agent checks the result against what you've asked for, and then it comes back to you.
Scene 2 of 5. Starting from what you say. From 0:54 to 1:39.
On screen: an example of the toolkit setting up a profile, from a demo recording. A page headed "Tell me what works for you", with a microphone button marked "Ready to listen".
Narrator: People told us early on that they didn't want to pick a category and just be filed underneath it. So the Ability profile agent starts from what you say, in your own words.
On screen: a label at the bottom right reads "Ability profile agent". It stays until the Adapter agent is named.
On screen: the microphone switches to "Listening", and the person's words appear on the page as they speak.
The person, in the demo recording: I use a screen reader. Ask me before anything is paid.
On screen: the toolkit replies under the request: "Got it. I updated your profile. I'll check with you before anything commits." A small note at the bottom says this is a research demo, with a synthesized voice and real speech recognition and toolkit response.
Narrator: You control what it records, and the app only sees what you allow. Say what you need once, and every app that uses it will follow.
On screen: as the narrator says this, a new line fades in under the reply: "Say what you need once. Every app that uses the toolkit follows it."
On screen: a second demo, headed "Your profile, an actual Wikipedia page" and "Set it once. Keep reading." The Wikipedia article about Lisbon fills most of the screen, in its usual small text. At the top right, the saved profile shows "Text scale 100 percent" and "Line spacing 1.5", with the line "The page is using the previous preference." A note under the page reads "Actual Wikipedia captures, saved extension profile, automatic page adaptation."
On screen: the label changes to "Adapter agent".
Narrator: In another example, focusing on the Adapter agent, another person might find that small text is hard to read, and they'll ask to turn on bigger text and spacing. The Adapter agent then changes the live page in front of them.
On screen: the profile changes to "Text scale 130 percent" and "Line spacing 1.8", and the line under the heading now reads "The page responds. No second setting on Wikipedia." The article redraws in larger text with more space between the lines.
Narrator: The site owner or developer can't make any of these changes. It has to come from the person's choice.
Scene 3 of 5. Seeing what changed. From 1:39 to 2:11.
On screen: an example of an AI agent booking a hotel, from a demo recording on a hotel booking site. Gray patches cover the site's name, its logo and its loyalty brand wherever they appear. A line at the top left of the screen says so: "Gray patches hide the booking site's branding." A label at the bottom right reads "Verification agent" for the rest of the scene.
Narrator: We also heard from people early on that they wanted to see what changed and stay on top of the decision. So that's where the Verification agent comes in. In this example, the person asks an AI agent to book a hotel on a hotel booking site.
On screen: a card headed "Task" with the person's request in quotes: "book a hotel in Lisbon for 3 people, 2 nights from 10 October 2026, free cancellation, under 200 euros a night". Below it, the booking site's search results for Lisbon, from 10 to 12 October, 3 adults and 1 room, filtered to 50 to 200 euros a night: "Lisbon: 4 properties found". A note under the page says the run starts from a search link with the city, dates, party size and budget already set.
On screen: the page for one hotel, with a "Reserve" button.
Narrator: The Verification agent reads each page as a screen reader would, and then it checks it against their request.
On screen: the hotel's table of rooms. A Triple Room shows 193 euros a night, 387 euros for 2 nights, with free cancellation. Under the price, in small print, a line is boxed: "Excluded: 6 percent VAT, 4 euros city tax per person per night." A printout line from the check appears below: "check: About 217 euros a night with the charges, over your 200."
The verifier voice, in the demo recording: About 217 euros a night with the charges, over your 200.
On screen: the hotel page scrolls to the "Reserve" button. A note under the page says the agent was held at a link named Reserve, lower on the page and out of view in this recording.
Narrator: In this case, the agent would stop and tell the person before they decide to move on.
Scene 4 of 5. The same profile across devices. From 2:11 to 3:13.
On screen: an example of one profile working across devices, from a demo recording headed "Your preferences travel with you". Three views sit side by side, each showing the same illustration of a lighthouse by the sea: a browser on the left, a headset in the middle and a phone on the right. Each view shows a caption, such as "Waves roll in. Seabirds call." In the headset, the caption floats under a framed picture in a gallery room, marked "world-locked". A note at the top right reads "Browser-hosted viewports. Working toolkit, hardware simulation."
Narrator: The same profile can actually carry beyond just a browser. Here, a person can ask for captions and less movement, and they choose what each device can read.
On screen: a label at the bottom right reads "Simulated headset and phone views".
Narrator: These are simulated headset and phone views, using working web pages that could be driven by the toolkit. Captions could switch on for a web video, or it could sit beside a presentation in headset view. In the phone view, they switch on too.
On screen: a bar along the bottom asks: "Use high-contrast captions everywhere? Suggested in the headset. Also increases display contrast. Nothing changed yet." It offers two buttons, "Keep current" and "Accept update".
Narrator: When the person asks for higher contrast, the change waits for their approval. If they agree, every view updates together.
On screen: the bar changes to "Accepted by you. Updated across all three apps", and in all three views the captions and the page switch to high contrast at once.
On screen: an example of an early demo app on smart glasses, from a second demo recording. A title card: "Setting up the ability profile on first use. Charging headphones." A label at the top left reads "Early demo app on smart glasses" until the end of the scene.
Narrator: There's already an early demo app using smart glasses.
On screen: an older adult wearing the smart glasses leans over a table with a pair of headphones and a charging cable on it. The recording's own subtitles show what is said. A panel at the top right, headed "Ability profile, setting up", changes as the volunteer speaks: pacing switches from "All at once" to "One step at a time", speaking rate slows from 1.0 times to 0.8 times, and clarity moves to "Extra clear".
The volunteer, in the demo recording: Tell me how to charge this, one step at a time, slowly, clearly, easy for anybody to understand.
On screen: the panel shows "Assistant thinking", then "Assistant speaking through glasses".
The glasses assistant, in the demo recording: Got it. You will connect the charging cable to the headphones and plug it in. Let me know when you have it.
Scene 5 of 5. Join in. From 3:13 to 3:37.
On screen: a made-here slide headed "Join in".
Narrator: This is early research, and we're sharing it before it's finished. If you'd like to take part, bring us a problem.
On screen: two cards. The first, "Bring a problem", reads "Share it through the contact details on the repository." The second, "Build something", reads "Point your coding agent at the repository, then tell us how it went."
Narrator: You can share it through the contact details on the repository, which is the AI for Accessibility Collective's toolkit on GitHub, linked here. Or you can point your coding agent at the repository and build something, then come back and tell us how it went.
On screen: the repository address, with the note "Contact details are on the repository". The address is github.com/AI-for-Accessibility-Collective/AI-for-Accessibility-Toolkit
End of the video.