Distributed LLM swarm. Web UI + Python orchestrator for running prompts across a bunch of peer-hosted local models (Ollama or any OpenAI-compatible endpoint).

Frontend: Chorus/frontend/ (Next.js). Backend: Chorus/orchestrator/ and Chorus/agent_backend/ (Python).

Click the button. On the Vercel import screen make sure Root Directory is set to Chorus/frontend. That field is pre-filled but worth double checking, since a wrong root is the usual reason you get a 404 after deploy.

Optionally set NEXT_PUBLIC_ORCHESTRATOR_BASE_URL to a hosted backend URL. If you leave it blank, users just paste a URL at /join and it gets cached in their session.

Almost always means the Vercel Root Directory is not pointing at the Next.js app.

1. Open the project in Vercel.
2. Settings > General > Root Directory > Edit.
3. Set it to Chorus/frontend and save.
4. Go to Deployments, pick the latest one, click Redeploy. Uncheck "use existing build cache".

Leave build command, install command and output directory at the Next.js defaults.

Full env var list is in Chorus/frontend/.env.example.

Vercel only hosts the frontend. You need to run the Python orchestrator yourself somewhere reachable over HTTP + WebSocket.

Code lives in Chorus/orchestrator/. Once it's up, either set NEXT_PUBLIC_ORCHESTRATOR_BASE_URL in Vercel or leave it empty and have users paste the URL at /join.

There's a Dockerfile and railway.toml tuned for Railway in the repo root.

1. New project > Deploy from GitHub repo, point it at this repo.

2. Railway picks up the Dockerfile. First build takes about 3 minutes.

3. Under Settings > Networking > Public Networking, generate a domain. Copy the URL (something like https://chorus-prod.up.railway.app).

4. Add a Volume under Storage with mount path /data. Without this, the SQLite job history and the Ed25519 signing key get wiped on every redeploy. Identity still works without a volume (ephemeral key on boot) but receipts won't be portable.

5. Set environment variables. The one you actually have to set for a hosted frontend to reach the backend is ORC_CORS_ORIGINS. Without it every browser request from Vercel gets blocked by CORS:

   Env var	Value	Required?
   ORC_CORS_ORIGINS	https://<your-vercel-domain> (comma-separated for multiple)	Yes, for hosted frontend
   CHORUS_DB_PATH	/data/chorus.db	Already set in Dockerfile
   ORC_KEY_PATH	/data/orchestrator_ed25519.key	Already set in Dockerfile
   ORC_LAN_MODE	0 for public deploys (default is 1)	Optional
   ORC_OPERATOR_TOKEN	Bearer token for operator endpoints	Optional

6. Back in Vercel, add the env var:

   NEXT_PUBLIC_ORCHESTRATOR_BASE_URL=https://<your-railway-domain>
   

   Apply to Production + Preview and redeploy. NEXT_PUBLIC_* vars bake in at build time, so you need the rebuild.

7. Go to /setup in your Vercel app. The wizard hits /health on the orchestrator before letting you finish, so if the URL or CORS is off, the error on screen tells you what's wrong.

Health check: GET https://<your-railway-domain>/health should return {"status":"ok"}. If it doesn't, check the Railway deploy logs.

Send new users to /setup in the deployed app. The wizard walks them through installing Ollama, pulling a model, and joining the swarm.

Chorus/
  frontend/        Next.js 16 + React 19 UI (Vercel)
  orchestrator/    Python signaling + broadcast server (self-host)
  agent_backend/   Python agent runtime