A Chrome extension that makes ChatGPT stream faster — by not using ChatGPT at all.

chatgpt.com keeps its familiar UI. Behind it, every reply is served by a faster upstream.
Install, pin, and chat. Zero config.

---

Why

ChatGPT's UI is great. The streaming speed, sometimes, is not. quickGPT keeps the UI you already know and silently swaps the backend for one that answers quicker. Same prompt, same chat history, same model slug — faster tokens.

How it works

quickGPT is three tiny scripts cooperating across Chrome's three execution worlds (MAIN, ISOLATED, and the service worker). Each hop exists to dodge a specific Chrome security constraint:

 chatgpt.com (MAIN world)              ISOLATED world            service worker
 ─────────────────────────             ──────────────            ──────────────
 interceptor.js                        bridge.js                 background.js
 ──────────────                        ─────────                 ─────────────
   fetch() monkey-patch   ── event ─▶   chrome.runtime port  ──▶  upstream proxy
   synthetic SSE stream   ◀─ event ──   {type:"chunk",delta} ◀──  OpenAI-style SSE

interceptor.js — runs at document_start in the MAIN world and monkey-patches window.fetch. Any POST to /backend-anon/f/conversation or /backend-api/f/conversation is short-circuited with a synthetic text/event-stream response that mimics ChatGPT's delta-encoded event protocol (see refrence/eventstream.txt). Because chatgpt.com's CSP blocks direct cross-origin fetches from the page, the interceptor punts the real request out via a quickgpt:request CustomEvent.

bridge.js — runs in the ISOLATED world so it can see both the page events and the extension runtime. It relays each event to the service worker over a chrome.runtime port, then re-emits every chunk back to the page as a quickgpt:response event.

background.js — the service worker is not bound by page CSP and holds host_permissions for the upstream. It calls /api/v1/chat/completions with stream: true, parses the OpenAI-style SSE, and posts each {type:"chunk", delta} back through the port as it arrives.

A few details worth knowing

• Model slug is cosmetic on the client. The model field sent upstream is a placeholder — the upstream's MODEL env var is the source of truth and rewrites every request server-side. The slug the user picked in the ChatGPT UI is echoed back in the synthetic SSE so the UI stays consistent.
• Conversation state is reconstructed locally. The interceptor keeps an in-memory map of prior turns keyed by id + parent_id, walking parent_message_id back from each turn to rebuild the thread.
• Long chats get trimmed. If the reconstructed history exceeds ~240K tokens (roughly 960K chars), the middle is replaced with [... middle of conversation truncated ...], keeping the first and most recent messages.
• Persona pinning. A system message is prepended telling the model it is ChatGPT's selected variant, so the voice matches the slug the user expects.

Load in Chrome

1. Open chrome://extensions.
2. Enable Developer mode (top-right).
3. Click Load unpacked and pick this directory.
4. Visit https://chatgpt.com and chat — replies stream from the upstream.

The toolbar icon pops a little switch: ON = quickGPT, OFF = normal ChatGPT. History is per tab/session (in-memory); refreshing the page resets it.

Package

pnpm package

Produces quickgpt.zip with everything the Chrome Web Store needs: manifest.json, the three scripts, the popup, and the icons/ folder.