Deploy and Host Open Intelligent UI on Railway

This template deploys Open Intelligent UI, an open-source implementation of the "Intelligent UI" pattern introduced in ChatGPT, built on OpenUI. Ask "I'm in San Francisco for a day, plan a sightseeing route" and the answer arrives as a map, photos and an itinerary you can edit, not a paragraph. One Next.js service, no database, and you bring your own model key.

About Hosting Open Intelligent UI

One service, web. Railpack reads engines.node from package.json and provisions Node 24, runs npm run build, then next start. There is no Dockerfile step and no image to pin: the build comes from the upstream repository on every deploy.

No database and no volume are attached. Conversations are stored in the browser's local storage, so a redeploy loses nothing on the server and there is no data to back up or migrate. Runtime calls leave the container for three things: your model endpoint, the public OSRM demo server for street routes, and Wikipedia/Wikimedia for photo search.

One model path is active at a time, chosen at boot by what is set: if LLM_BASE_URL has a value, the app talks to that OpenAI-compatible endpoint; otherwise, if MODEL_PROVIDER=ollama or OLLAMA_MODEL is set, it talks to Ollama; otherwise it falls back to the OpenUI Gateway at https://api.thesys.dev/v1/embed. Both path selectors ship empty, so the gateway path is what runs until you set LLM_BASE_URL. The two keys are alternatives: set THESYS_API_KEY, or set LLM_BASE_URL and LLM_API_KEY together, never both. If neither is set the app says so on the first message rather than failing silently.

Why Deploy Open Intelligent UI

Running it locally means Node 24, an install of roughly 400 packages, and a key. Deploying gives you a URL you can open on a phone or hand to someone else, with the model key held server-side instead of pasted into a browser session. It is also the shortest way to see whether streaming UI instead of streaming text is worth building into your own product, before you commit to a component library.

Common Use Cases

Try generative UI hands-on, from a URL, without a local toolchain
Demo "answers that render as interface" to a client or a team
A starting point for a vertical app: keep the streaming loop, replace the component library
Compare how different models lay out the same answer, since the layout is written by the model at runtime
A sandbox for prompt and tool changes against the OpenUI spec

Dependencies for Open Intelligent UI Hosting

An OpenAI-compatible model endpoint, or an OpenUI Cloud key. Nothing else is required.
No database, no object storage, no volume.
Outbound network access to your model endpoint, router.project-osrm.org, and the Wikimedia APIs.

Deployment Dependencies

Upstream repository (MIT)
OpenUI, the generative UI standard the app renders
Railway healthchecks
Railway variables

Implementation Details

Variables

The two API keys are alternatives, never both. Which one matters depends on LLM_BASE_URL: while it has a value, the direct path runs and only LLM_* is read; clear it and the gateway path runs, and only THESYS_* is read. Yours to choose at deploy time.

Variable	Path	Required	Default	Purpose
`LLM_BASE_URL`	direct	path selector	empty	Any OpenAI-compatible base URL ending in `/v1`, e.g. `https://openrouter.ai/api/v1`. Setting it switches the app off the gateway. Leave empty to use `THESYS_API_KEY`.
`LLM_API_KEY`	direct	yes, with that path	—	Key for `LLM_BASE_URL`. OpenRouter keys come from `openrouter.ai/keys`. Ignored while `LLM_BASE_URL` is empty.
`LLM_MODEL`	direct	yes, with that path	empty	Model id the endpoint serves, set together with both values above, e.g. `openai/gpt-6.1-sol` on OpenRouter. A model the endpoint does not have fails on the first prompt.
`LLM_REASONING_EFFORT`	both	no	`low`	`low`, `medium`, `high`, or `default` to send no reasoning parameter at all. Use `default` if a model rejects the parameter.
`THESYS_API_KEY`	gateway	yes, with that path	—	OpenUI Cloud key from `thesys.dev`. This is the path that runs while `LLM_BASE_URL` is empty.
`THESYS_MODEL`	gateway	no	empty	Model id on the OpenUI Gateway. The app falls back to `openai/gpt-6.1-sol` when this is empty. Ignored while `LLM_BASE_URL` is set.

First use: open the service URL and ask for something with structure. The first answer takes a few seconds while the model writes the UI program. If chat returns an error, check which path is active first: with LLM_BASE_URL set, a missing or wrong LLM_API_KEY is the usual cause; with it empty, THESYS_API_KEY is.

Limits worth knowing: chat is light theme only, the dark theme lives behind a dev-only route upstream. Layout quality tracks the model; smaller models produce simpler cards. The OSRM and Wikimedia calls hit public demo endpoints with rate limits outside this template's control.

Notes: this template is published by a third party. It is not affiliated with the upstream maintainers, OpenUI, or OpenAI. Upstream issues belong in the upstream repo; template issues (variables, build, networking) belong in the template queue.