How LuisCode actually works, end to end -- what's real, what's generated, and what runs where.
LuisCode is an AI builder backed by DeepSeek. You describe what you want, and it writes real files -- not a template fill-in. Everything it generates is genuinely tested: CLI tools are actually run on a real Cloud machine before being handed back, and hosted sites are actually served from that same machine over a real tunnel.
There's no login required to use the builder or Agent Chat -- both are open to everyone. Signing in with a LuisSearch account is optional and mainly unlocks web search running as your own account instead of anonymously (see Account & sign-in).
Pick one when you describe a build:
| Type | What you get |
|---|---|
| Website | Static HTML/CSS/JS, optionally hosted live with one click. |
| Desktop | A site framed and described as a desktop-style app (still HTML/CSS/JS under the hood). |
| Mobile | Same, tuned for a phone-sized layout. |
| CLI | A real script (the AI picks the language). It's actually executed once on the Cloud machine to confirm it runs, including installing whatever real dependencies it needs. |
| Everything | Website + desktop + mobile + CLI variants generated together in one request. |
Check any combination of these to have the generated site call a real external API instead of using fake data. You can select more than one at once.
LuisSearch · DuckDuckGo · Wikipedia · Weather · GitHub · Countries · Trivia · Dictionary · Advice · Books · Dogs · Cats · Holidays · Crypto · Jokes
Each one wires the AI's system prompt with the real endpoint shape for that API, so the generated JS calls it correctly on the first try.
A ChatGPT-style chat mode, but the assistant has a real run_command tool with no timeout and no sandboxing beyond the tool itself -- it genuinely installs packages, writes files, and runs processes on a dedicated Cloud machine. It can also host whatever it builds publicly with host_project, and search the web.
Responses stream in token-by-token as the model generates them, and every real command it runs shows up as a step card you can expand to see the exact command, output, and exit code.
The main agent can delegate a focused sub-task to a subagent via spawn_subagent, so a long investigation or build step doesn't clutter the main conversation with every intermediate command. Two roles:
run_command. It's blocked from installing, deleting, or overwriting anything; good for figuring out which package/API/approach to use.run_command/host_project/web search as the main agent, for a sub-task that needs to actually build something.A subagent's own trace shows up nested under its parent step card.
The agent's web_search tool is backed by LuisSearch -- a real, independently built search engine, not a third-party API. If you're signed in (see Account), searches run as your account rather than anonymously.
LuisCode Desktop is a native SwiftUI app -- a real code editor (file tree, monospaced editor) plus the same Agent Chat, built via a private GitHub Actions pipeline rather than requiring Xcode locally. First launch walks through a short setup: welcome, sign in with LuisSearch, real backend health checks, then you're in.
Opening a folder automatically zips and uploads it to the Cloud machine, so the agent can immediately build/run against your real project files.
A panel in the desktop app that shows what the agent has actually created on the Cloud machine (under its project directories), refreshed automatically after every agent turn. Click any file to view its contents without leaving the app.
LuisCode has no separate account system of its own -- it reuses your LuisSearch account. Signing in is entirely optional; the builder, hosting, and Agent Chat all work anonymously. The one thing it unlocks is web search running under your own LuisSearch account/rate tier instead of anonymously.
Everything on this page is genuinely real, running infrastructure -- not mocked or simulated for demo purposes. A few things worth knowing as a user: Generation and Agent Chat are rate-limited per hour to keep the service available for everyone. CLI/Everything builds take real time because each one is genuinely executed and verified on the Cloud machine, not instantly templated. Hosted sites and Agent Chat tunnels can occasionally roll over to a new address as part of normal operation. If something looks unreachable, check back shortly -- it's usually a routing refresh, not a bug in what you asked for.