Docs Demo Dashboard Get your key
On this page
  1. How it works
  2. Making an AI layer
  3. Claude Code
  4. Codex, GitHub Copilot and Cursor
  5. What the pins say
  6. Talk instead of typing
  7. AI layers or the connector?

AI layers

An AI layer is a private conversation between you and the coding AI working in your project. Point at something on your running app and say what you want; the comment goes straight to the AI, which changes the code and answers in the thread.

How it works

AI layers are part of the Pro plan.

Making an AI layer

  1. Open your app on localhost in Chrome, with the extension or with the SharedLayers script your app loads, and choose its project folder if you have not yet.
  2. In the layer menu choose New layer, then AI, then Next.
  3. Pick your AI app and connect it, using the steps for it below. The setup waits until it sees the app.
  4. Confirm the project, and create the layer.

The same steps are the layer's Settings, in its ⋯ menu, for changing the app or the project later. While the AI is not connected, a Connect button sits beside the layer's name.

Claude Code

Install the SharedLayers plugin once. In Claude Code, run:

/plugin marketplace add SharedLayers/claude-plugin
/plugin install sharedlayers@sharedlayers

From then on, every Claude Code session you start in the project listens for the layer's comments by itself, in the terminal and in the desktop app, and is woken by a new one even when it is idle. Nothing else to start.

Claude Code will ask before it edits files unless you have allowed that. Each question pauses the request until you answer it in Claude Code.

Codex, GitHub Copilot and Cursor

Start SharedLayers for your app in a terminal in the project folder, and leave it running while you work. It hands each comment to the app's command line, one at a time:

AppRun in the project
Codexcurl -fsSL https://sharedlayers.com/ai.mjs | node --input-type=module - codex
GitHub Copilotcurl -fsSL https://sharedlayers.com/ai.mjs | node --input-type=module - copilot
Cursorcurl -fsSL https://sharedlayers.com/ai.mjs | node --input-type=module - cursor

Once the project has its .sharedlayers folder, the same is node .sharedlayers/ai.mjs codex (or copilot, cursor). Each needs that app's command line installed and signed in first:

AppInstallSign in
Codexnpm i -g @openai/codexcodex login
GitHub Copilotnpm i -g @github/copilotcopilot login
Cursorcurl https://cursor.com/install -fsS | bashcursor-agent login

Each runs on your own plan with that app.

What the pins say

A pin on an AI layer shows your face with the AI app's mark, and a word on what is happening:

LabelMeans
SentWaiting for the AI to take it. Queued #2 when others are ahead of it.
WorkingThe AI is on it, with a light running round the pin. Where the app says, the label says what it is doing: “Editing styles.css”.
QuestionThe AI asked you something in the thread. Reply to carry on.
A tickDone: the AI answered with what it changed. Unread until you open it.
No code connectedThe AI app is not listening. The comment waits, and goes as soon as it is.

Talk instead of typing

On an AI layer the toolbar has a Talk button, in the extension and in the SharedLayers script on an app on your computer: it turns on the microphone and a pen. Say what you want while you circle or point at things on the page, and each sentence becomes a comment on the element you drew on. Your AI turns what you said into a clear request when it takes it; the thread keeps what was heard. Holding Alt (⌥) does the same for as long as you hold it. Talk is part of Pro, in Chrome.

AI layers or the connector?

Both let an AI work on comments. The connector is for comments other people leave, on any site and any layer: you ask your AI app to go through them. An AI layer is for you and your AI alone, on your own computer, with each comment sent as you write it.