CarradosInsight Without Sight

The Making of gent

Charli-Jo Tyrer ·

This afternoon I typed one line into Claude Code:

can you look at nvda repos on github

By the evening, gent was running inside my copy of NVDA.

I press NVDA+G. I type "way too slow, speed it right up". NVDA speeds up. I type "undo that". It goes back. I type "describe this image". It politely tells me it can't do that yet.

That sounds like a vibe-coding story. It isn't one, and the difference is the point of this post.

Vibe coding and agentic coding

Vibe coding is when you describe what you want, an AI writes something that looks plausible, and you run it and hope. If it breaks, you paste the error back and hope again. Nobody reads the code. Nobody checks the platform it targets. It's fine for a toy.

It is not fine for a screen reader. For the people who rely on NVDA, it is how they use their computer. An add-on that crashes it, silently changes the wrong setting, or quietly sends what I type somewhere unexpected is not a toy. It's a locked door.

Agentic coding is different. The agent doesn't just write code. It works the way a careful developer works: it reads the real source, checks its assumptions, runs tests, deploys infrastructure, finds problems in the real world, and stops to ask me when a decision is mine.

Here is what that looked like.

It read NVDA before it wrote for NVDA

The first thing Claude did was look at what NV Access actually publishes on GitHub: the screen reader itself, the add-on store, and the official add-on template.

When I said I only wanted to write add-ons, it set up a local workspace from that official template, wrote small scripts to create, build and install add-ons, and built a throwaway test add-on to prove the pipeline worked before deleting it.

Even that test caught something. The starter command it generated gave the test add-on the key NVDA+Shift+S, which is NVDA's own sleep mode. Claude changed the template so new add-ons don't claim any key by default.

Later, before writing a line of gent, it read NVDA's source code for my exact version, 2026.2, not the development version:

That mattered. "Key echo" turns out not to be on or off. NVDA stores three values: off, only in edit controls, and always. Punctuation is a scale with its own numbers. An add-on written from memory would have got both wrong.

It also checked that NVDA+G isn't already used by NVDA. (I first asked for NVDA+J, then changed my mind halfway through its turn. It picked that up.)

It fact-checked me

Before any code, I gave Claude two Blind Ant design-fiction pieces, NVDA Commander and NVDA gent, and asked what it thought.

It checked their claims against TypeSafe's documentation, OpenAI's developer docs and NVDA's user guide. Most held up. One didn't: the pieces estimated that £50 of inference would buy "tens of thousands" of decisions. At Jev's published price, it's closer to two million. That was out by about a hundredfold, and it changes what the "Jevons test" would actually measure.

It also pointed out that an AI image-description add-on for NVDA, AI Content Describer, already exists, and that Sign in with ChatGPT is a preview limited to paid plans. That's why the first version of gent uses Jev alone.

A vibe coder agrees with you. An agent with access to the sources sometimes tells you you're wrong.

A real architecture, not a script

gent has three parts.

The add-on, in Python, inside NVDA. It opens a text box when you press NVDA+G, sends what you typed, then makes the change itself using ordinary NVDA code. It remembers its last five changes so it can undo them.

The relay, a Cloudflare Worker. It sits between the add-on and the AI so that my TypeSafe key never ships inside the add-on. It also:

Jev, TypeSafe's decision model. It never writes text. It is given the list of things gent can do, and what the user typed, and it returns one choice with a probability. It cannot invent a setting that isn't on the list.

The add-on and the relay are versioned. When I asked for a help overview, the relay learned to offer "help" only to add-ons that understand it. An older copy of gent can never be handed an action it doesn't know.

The secret key stayed mine

At one point I said: the API key is on my clipboard.

Claude wouldn't touch it. It said keys are mine to enter, deployed the relay without one, and told me exactly where in the Cloudflare dashboard to paste it as an encrypted secret. Then it confirmed the secret was there without ever seeing it.

That's a small moment, and it's the clearest difference from vibe coding I can think of. The agent knew where its job ended.

It found problems in the real world

Tests on my own machine only prove so much. Once the relay was live, real problems appeared, and Claude found them by testing against the real thing.

Fourteen real phrases went through the live system before I installed anything. Jev got all fourteen right, in 0.3 to 0.6 seconds each, for a tiny fraction of a cent in total.

Tests, style checks, commits

Every change came with checks:

Everything went into git, committed under my own account with an honest note that Claude co-wrote it. I chose to work on the main branch. Claude asked first.

And it's all public. Both the add-on and the relay are open source, under the same GNU General Public License as NVDA's add-on template, on GitHub at Carrados/nvda-addons. Nothing secret is in there. The key lives only in Cloudflare.

What I actually did

I didn't write the Python, the JavaScript or the Cloudflare configuration.

I did:

That's the shape of agentic coding, at least as I've now experienced it. The person decides. The agent does the engineering, checks its work against reality, and hands back the decisions that belong to the person.

From a prompt to an add-on

gent is still a design-fiction experiment. It changes three kinds of setting. It has been tested by exactly one person, me, in one afternoon.

But it is real. It reads NVDA's actual settings. It runs through actual infrastructure with actual limits. It keeps an actual secret. And it was built properly, not hoped into existence.

Type "help" and it will tell you what it can do.

It waits to be asked.

Try it, read it, fork it

If you try it, tell me what you asked it that it couldn't do. That's the most useful thing this experiment can learn.