LeafLLM

LeafLLM is a privacy-respecting browser extension that brings the power of large-language models (LLMs) into your browser. Right-click any text and choose Summarize, or use the Complete / Improve / Ask keyboard shortcuts inside an editor (e.g. the Overleaf source editor). It works with any OpenAI-compatible Chat Completion API — point it at OpenAI, a local vLLM / llama.cpp server, or any other compatible endpoint by setting the Endpoint URL, API key and model in the settings popup.

The extension originated from GPT4Overleaf. This fork is cross-browser: it works in both Mozilla Firefox and Google Chrome. The code uses the browser.* WebExtensions API and a WebExtension Manifest V2, both of which are supported by Firefox. The keyboard-shortcut tools (recall the original GPT4Overleaf Alt+C / Alt+I / Alt+A) are implemented and described in Usage.

Installation

Firefox (temporary load — for development / testing)

1. Clone this repository.
2. Open Firefox and navigate to about:debugging#/runtime/this-firefox.
3. Click "Load Temporary Add-on...".
4. Select the manifest.json file in the repository folder.
5. The LeafLLM icon appears in the toolbar. (Temporary add-ons are removed when Firefox restarts.)

Firefox (permanent / distribution)

To install permanently, the add-on must be signed through addons.mozilla.org (AMO). You can build and sign a package with the web-ext CLI:

npm install -g web-ext
web-ext build                      # creates a .zip in web-ext-artifacts/
web-ext sign --api-key=<JWT_ISSUER> --api-secret=<JWT_SECRET>

The browser_specific_settings.gecko.id (leafllm@bthink.bgu.ac.il) in manifest.json is required for signing and is already set.

Chrome (manual installation)

1. Clone the repository.
2. Open Chrome and go to chrome://extensions/.
3. Enable developer mode.
4. Click "Load unpacked" and select the repository folder. (Since the code uses the browser.* namespace, Chrome needs the small webextension-polyfill shim, or you can replace browser. with chrome. — note that Chrome's chrome.* callback APIs differ from the promise-based browser.* API.)

Configuration

Click the LeafLLM button in the browser toolbar to open the settings popup. There are three fields, all stored locally using the browser's local storage (so changing them takes effect immediately on the next invocation):

• API Key — the bearer token sent to your endpoint (e.g. an OpenAI API key).
• Endpoint URL — the Chat Completion endpoint to call. Defaults to https://api.openai.com/v1/chat/completions if left blank. Point this at any OpenAI-compatible server (vLLM, llama.cpp, LM Studio, LocalAI, etc.).
• Model — the model name to pass in the request body (e.g. gpt-3.5-turbo, gpt-4o-mini, or the name your local server exposes). Defaults to gpt-3.5-turbo if left blank.

Test connection

The Test connection button sends a minimal request to your endpoint and reports the result. If your Endpoint URL doesn't accept POST requests (a common mistake is pasting only the server's base URL — e.g. https://host instead of https://host/api/chat/completions), it automatically tries a few common OpenAI-compatible paths appended to that origin:

• /v1/chat/completions
• /api/chat/completions
• /openai/chat/completions
• /api/v1/chat/completions

If one of them works (or at least accepts POST), the popup asks you to confirm switching to it. For an OpenWebUI instance the correct path is typically /api/chat/completions. The request uses max_tokens: 1 to keep the test cheap, and a list of every path tried (with its HTTP status) is shown in the result area.

Usage

Summarize (context menu)

1. Select any text on a web page.
2. Right-click and choose Summarize (the LeafLLM context-menu item).
3. A small popup appears in the top-right corner of the page showing a loading indicator, then the summary. Click the × to dismiss it.

Editor shortcut tools (Overleaf and any editable area)

Select text inside an editor (e.g. the Overleaf source editor, which lives in an iframe, or any other text area / contenteditable element) and press:

• Alt+C — Complete: continues the text after your selection (the selection is kept, the continuation is appended right after it).
• Alt+I — Improve: comments out the selected text (% … per line) and inserts the improved text below it.
• Alt+A — Ask: replaces the selected text with the model's answer.

A small status popup appears in the top-right while the request is in flight; this works inside the editor frame (the content script runs in all frames). If the editor cannot be edited programmatically, the result is shown in the popup instead. The shortcuts only fire when you have a selection inside an editable element, so they won't interfere with normal Alt-key usage elsewhere.

Note on Ask: it sends the selected text as a user instruction and replaces the selection with the model's answer, so it's meant for short instructions (e.g. "Create a table 4x3 with a bold first row"), not for re-explaining whole sections. Requests abort after 60 s; if a request takes longer (slow model, long generation, or a proxy that stalls long POSTs) you'll get a clear "Request timed out" error instead of a spinner that never stops.

How it works

• scripts/main.js — background page; registers the Summarize context menu, reads the API key, endpoint URL and model from storage.local, and POSTs the selected text to the OpenAI-compatible Chat Completion endpoint via fetch. Also handles run-tool messages from the content script for the complete/improve/ask shortcuts (each tool uses a different system prompt) and delivers results back to the originating frame with tabs.sendMessage using sender.frameId (the Overleaf editor lives in an iframe). Requests use stream: false and a 60 s timeout.
• scripts/content-script.js — runs in all frames. Shows the summarize loading/result popup (top frame only); listens for the Alt+C / Alt+I / Alt+A shortcuts in editable areas, asks the background to call the LLM, and inserts the result back into the editor via document.execCommand.
• popup/popup.html / popup.js / popup.css — settings UI for the API key, endpoint URL and model.

The request sent to the endpoint is a standard Chat Completion call:

{
  "model": "<model>",
  "messages": [
    { "role": "system", "content": "<tool-specific prompt>" },
    { "role": "user", "content": "<selected text>" }
  ]
}

The system prompt differs per tool:

Tool	System prompt
Summarize	"Summarize this for a second-grade student:"
Complete	"You are an assistant in a Latex editor that continues the given text. No need to rewrite the given text."
Improve	"You are an assistant in a Latex editor that improves the given text."
Ask	"You are an assistant in a Latex editor. Answer questions without introduction/explanations."

For Summarize, the first choice's message.content is shown in the page popup. For the editor tools, it is inserted directly into the editor.

Privacy

The extension saves its configuration (API key, endpoint URL and model) locally on the user's computer using the browser's local storage (it does not sync across devices). The API key and the selected text are sent only to the endpoint URL you configure, and only for the purpose of generating a response (summarize / complete / improve / ask). The extension's authors are not responsible for what the chosen endpoint provider does with this data. The authors do not collect any data from the extension's users.

Firefox compatibility notes

This fork was adapted from the original Chrome extension to run cleanly in Firefox. The changes are minimal because the extension already used the cross-browser browser.* API:

• Icons: fixed icon paths to the bundled images in popup/ (the original manifest pointed at a non-existent icons/ folder).
• Host permission: uses <all_urls> in permissions so cross-origin fetch works to any user-configured OpenAI-compatible endpoint (the endpoint URL is set at runtime, so a single broad host permission is required).
• CSP: simplified content_security_policy to script-src 'self'; object-src 'self'; (removed an unused https://esm.run remote source).
• browser_specific_settings.gecko.strict_min_version: set to 115.0 to declare a supported Firefox baseline.
• Provider switch: replaced the bundled Google Generative AI (Gemini) SDK with a direct, dependency-free fetch to an OpenAI-compatible Chat Completion endpoint; removed the now-unused scripts/gemini.js and scripts/lib/generative-ai.js.
• Storage: switched from storage.sync to storage.local. storage.sync requires a Firefox account to sync and is unreliable for unsigned/temporary add-ons, which previously caused settings changes (e.g. the endpoint URL) not to take effect.
• Editor tools: re-implemented the Alt+C / Alt+I / Alt+A shortcut tools. The content script runs with all_frames: true so the shortcuts work inside the Overleaf source editor's iframe, and results are inserted via document.execCommand('insertText') (works with the CodeMirror editor).

Ask examples

For example, select the text "Create a table 4x3 that the first row is boldface" and press Alt+A; it will be replaced with, e.g.:

\begin{tabular}{|c|c|c|}
\hline
\textbf{Column 1} & \textbf{Column 2} & \textbf{Column 3}\\
\hline
Entry 1 & Entry 2 & Entry 3\\
\hline
Entry 4 & Entry 5 & Entry 6\\
\hline
Entry 7 & Entry 8 & Entry 9\\
\hline
\end{tabular}

You can then, for example:

1. Write before the table: "Place the following tabular inside a table environment, center it, and give the following title: The comparison of the three approaches"
2. Select that sentence and the table
3. Press Alt+A to trigger the ask tool.

The result will be:

\begin{table}[h]
\centering
\caption{The comparison of the three approaches}
\begin{tabular}{|c|c|c|}
\hline
\textbf{Column 1} & \textbf{Column 2} & \textbf{Column 3}\\
\hline
Entry 1 & Entry 2 & Entry 3\\
\hline
Entry 4 & Entry 5 & Entry 6\\
\hline
Entry 7 & Entry 8 & Entry 9\\
\hline
\end{tabular}
\end{table}

Future work

Possible improvements to consider:

• Advanced per-tool customization (temperature, max_tokens, custom system prompts) exposed in the popup, similar to the old JSON-configuration screen.
• Streaming responses for the editor tools.
• Treating plain <textarea>/<input> editors (which ignore execCommand('insertText') in some browsers) with a dedicated insertion path.

Issues

If nothing happens when you use the plugin, verify that the extension's shortcuts are not in conflict with other extensions' shortcuts. In Chrome, go to chrome://extensions/shortcuts.

If you encounter any problem or question, please open an issue in the project's repository.