less-tokens — Prompt Compressor for VS Code & Cursor

A small composer panel. Write a prompt (or build zones), pick which compression techniques to apply, optionally attach an image / PDF / Word file (OCR or text extraction), Compress, then Insert the result into Copilot / Cursor chat. It’s just the prompt compressor — no second chat, no model calls.

What it can and can’t do

It cannot intercept the requests Copilot or Cursor send to their own servers — those are closed, and no extension can rewrite them. "Insert into chat" copies the compressed prompt and prefills the native chat input where the editor allows it (VS Code’s built-in Chat); in Cursor you paste with Ctrl/Cmd+V. You always review before sending.

What it shrinks is the context you write — which is where the avoidable token spend lives.

Features

• Normal mode — a single prompt box.
• Structured mode — stack free (full compression) / careful (safe only) / protected (untouched, for schemas & formats) zones; add as many as you like.
• Techniques — toggle any of the eleven less-tokens techniques.
• Attach a file — image → OCR (reduce_image), PDF/Word → text extraction (reduce_document); optional "compress further" on the extracted text.
• Live token estimate and a compressed preview before you insert.

Requirements

A reachable less-tokens FastAPI backend (the same one the web tester uses). Set lessTokens.apiUrl to it — http://localhost:8000 locally, or your deployed URL. The backend must expose /smart_compress_batch, /compress_structured, /reduce_document, and (for image OCR) /reduce_image.

Install

No build step — plain JavaScript.

Try it from source: open this folder in VS Code/Cursor and press F5.

Install permanently:

cd less-tokens-vscode
npm install
npm run package          # produces less-tokens-0.2.0.vsix

Then: Command Palette → Extensions: Install from VSIX… → pick the file.

Use

• Ctrl/Cmd+Alt+K — open the Prompt Compressor (or click compress in the status bar, or run it from the Command Palette).
• Compose → Compress (preview + savings) → Insert into chat ▸.
• Bonus: select text in any file → Ctrl/Cmd+Alt+L compresses it straight to the clipboard, no panel.

Settings

• lessTokens.apiUrl — backend base URL.
• lessTokens.flags — default technique toggles (the panel’s toggles start here).
• lessTokens.showStatusBar — show/hide the status-bar button.

Notes

• Token numbers are estimates (~4 chars/token); the real count is whatever the model reports.
• Compression is lossy on prose by design; code, tables, URLs and math are preserved by smart_compress. Read the preview before sending.
• Image OCR needs RapidOCR installed in the backend; if it isn’t, the attach step reports the backend’s error instead of failing silently.