Self-hosting LibreChat: one ChatGPT UI for your whole team, billed per token instead of per seat

A developer detailed how to self-host LibreChat, an open-source ChatGPT-style interface, to replace per-seat subscriptions with per-token API billing for teams. The guide covers setup, security fixes, and common failures, noting that a five-person team could save from $100/month on seats to about $20-40 in hosting and API costs. The developer also disclosed co-founding livemy.app, which offers a one-click LibreChat template.

ChatGPT Plus is $20 per person per month. A team of five is $100/month, forever, whether anyone used it much or not. LibreChat is the open-source ChatGPT-style interface that turns that per-seat subscription into a per-token API bill. One instance, one login page, and whichever models you plug in behind it: OpenAI, Anthropic, Google, Azure OpenAI, Groq, DeepSeek, or a local model through Ollama. This is the setup, the security pass most quick-starts skip, and the five failures that show up once real people use it. Current LibreChat v0.8.x is well past "a chat window." It ships agents, MCP support, a code interpreter, artifacts, file chat and proper multi-user authentication. The feature list reads like a paid product because it is chasing several of them at once. The part that matters architecturally: MongoDB holds accounts and conversation history . That is where your team's accumulated context lives, and it is the thing you can lose. git clone https://github.com/danny-avila/LibreChat.git cd LibreChat cp .env.example .env docker compose up -d Open port 3080. Register. The first account created becomes the admin — do this yourself before anyone else has the URL. That is the honest quick-start. What it hides is everything around it: a reverse proxy and TLS so it is not served over plain HTTP, keeping the MongoDB volume persistent, updates LibreChat ships often , and backups. Budget 1–2 hours for a proper first deploy, not fifteen minutes. Disclosure so you can weigh the rest: I co-founded livemy.app https://livemy.app , which has a one-click LibreChat template. Coolify and Dokploy also ship LibreChat templates and are the obvious picks if you already run your own VPS, and plain Compose on a $5 box is the cheapest path if you enjoy servers. A flat subscription and a token bill charge for different things. $20/month flat is a bargain if you use it constantly, and mostly a round number if you use it a few times a week. API pricing bills for exactly the tokens the model read and wrote, at rates the providers set directly. For light-to-moderate users, which is most of a team, API usage lands at a fraction of a flat seat. Running costs: Five-person team: ~$100/month on seats versus roughly $10 hosting plus $10–30 of actual API usage. The gap grows with every seat you do not buy. The honest counter-case: one heavy solo user, all day, every day. Flat $20 can beat token pricing there. LibreChat is a team play. LibreChat without keys is a beautiful empty shell. The app runs, the model list is blank or every message errors. It does not come with AI, it comes with sockets for AI. Fix. Create keys with at least one provider, set them in your environment, restart, send one test message before inviting anyone. .env The example config ships placeholder JWT secrets. Leave them and you share session-signing keys with every other instance that also skipped this step, which defeats the point of having authentication at all. Fix. Generate long random values before first launch: openssl rand -hex 32 By default, anyone who finds your instance can create an account, and every message they send spends your API budget. A public LibreChat with open signup is free ChatGPT for strangers, billed to your card. Fix. Disable open registration once your team is in. Set provider-side spending caps as the backstop, because config is a policy and a cap is a wall. All history, accounts and settings live in MongoDB. If the volume is not persistent, the next container rebuild wipes everyone's conversations. For a tool whose entire value is accumulated context, this is the most demoralising possible failure. Fix. Verify persistence before real use: restart the whole stack, confirm chats survive. Then schedule backups. Core LibreChat is light. Enable search indexing and file-chat RAG and memory climbs. On a small instance that shows up as random slowness and OOM-killed containers, not a clear error message. Fix. 2 GB for chat-only. Enabling search and file chat for a team means more, or turn features on one at a time and watch memory. Before you hand out the URL: Five minutes. Skipping step 3 or 4 is how people end up funding a stranger's model usage. LibreChat is the rare self-hosted app where the economics argue as loudly as the ideology. The per-seat model does not survive contact with a shared instance and API keys. If you are one person using AI all day, keep your subscription. If you are a team of three or more where usage is uneven, this pays for itself the first month. What did you hit first when self-hosting it? I am betting on either the empty model list or the wiped Mongo volume, and I would like to know which one wins.