A free AI that vibe codes in your browser.
CheapGPT writes it, runs it, breaks it, then ships it anyway — a coding AI with persistent-ish memory that builds real apps live in a browser tab. No setup, no login, no refunds (it’s free).
What CheapGPT is
CheapGPT is a free-to-use AI that writes and runs code: a chat that remembers you, and a coding agent that builds working apps in the browser and ships them to your own GitHub and Railway. It comes with persistent memory, an agentic build loop, and real tools for reading and writing your project.
There’s no account and no paywall. Identity is an anonymous device_id in your browser. The bill is meant to be paid by a crypto token’s buy/sell tax, so people pay nothing — which makes the two design rules non-negotiable: keep it simple, keep it cheap. The agent runs in a WebContainer (Node in your own browser), so heavy compute stays $0.
How it works
Three moving parts — chat, memory, and the agent — over one anonymous identity. The server never touches a filesystem, mostly so it has nothing to blame.
Chat
A stable prompt prefix keeps CheapGPT’s cache warm, streaming the first token in about 1.4s. Every call sends data_collection: deny — your chats are never used for training.
Memory
One plain-text rolling summary per device — no vectors, no RAG, no long-term memory to speak of. Rewritten by the cheap model every 10 turns, plus a “knowledge” block it swears it’ll never forget.
The agent
CheapGPT streams its thinking and calls tools; the browser runs each one in a WebContainer, feeds the result back, and loops (up to 12 steps, then resumable). You get a live preview iframe and a one-click .zip.
Ship it
Push the build to your own GitHub via the OAuth device flow (token stays in your browser, never on our servers), then deploy to your own Railway — in-app with a token, or by deep-link. Your account, your bill.
What you can do
Streaming chat
Talk to CheapGPT with persistent memory across sessions. ~1.4s to first token.
Coding agent
Writes files, runs a dev server, fixes its own errors, hands back a live preview — all in your browser.
Templates
Start from a landing page, REST API, Telegram bot, scraper, React app, or “explain a repo”.
Ship to GitHub
One click or a chat command creates a repo on your account and pushes the build in one commit.
Deploy on Railway
Deploy to your own Railway account in-app, with an env-var editor. Deep-link path too.
Share a build
Publish any build to a public read-only link — a showcase page with the code and a “build your own” CTA.
Edit & re-run
Tweak any file in the Code tab and save — it writes to the WebContainer and reloads the preview.
Knowledge & memory
Set custom instructions once; CheapGPT always sees them. Memory keeps the thread across sessions.
Image & video
Generation is built in and gated behind a key — flips on when the treasury is ready.
SoonThe stack
Deliberately small. Every piece chosen to stay simple, cheap, and hard to sue.
API routes
Thin server routes — mostly proxies and persistence. The heavy lifting (and the blame) happens client-side.
| Route | Method | Purpose |
|---|---|---|
| /api/chat | POST | Streaming chat (deep:true = deep mode) |
| /api/agent | POST | One agent step, streams NDJSON (deltas + tool calls) |
| /api/projects | GET · POST | Save / list / load agent project file trees |
| /api/share | GET · POST | Publish a build; public read of a shared project |
| /api/github | POST | GitHub device-flow proxy (no secret stored) |
| /api/railway | POST | Railway GraphQL proxy (token pass-through) |
| /api/memory | GET | Read the rolling summary |
| /api/knowledge | GET · POST | User custom instructions |
| /api/media | POST · GET | Image / video via Fal (off until keyed) |
| /api/usage | GET | Operator cost summary |
| /api/health | GET | Config + which keys are wired |
Free, on purpose
Cost control isn’t a feature here — it is the entire business plan. Compute lives on your machine; CheapGPT is cheap to run; a token tax covers the rest. Somehow.
Roadmap
Shipped is live today. Next is queued. Later is architecture-level and weighed against “keep it cheap.”
Shipped
- Streaming chat + memory
- Coding agent + live preview
- Templates
- Ship to GitHub
- Deploy on Railway + env vars
- Share a build
- In-preview file editing
Next
- Token-gated GPT-4o & o1 access
- Fork / remix a shared build
- Static-host deploy (Vercel / Netlify)
- Agent memory of past builds
- Streaming diff view
- Image & video generation on
- Lock the brand
Later
- Cloud terminal (token-funded)
- The agent’s own browser
- Accounts & cross-device sync