All The AIs

What it was for

By 2024 the best multi-model AI interfaces were open source and, for most people, entirely out of reach.

LibreChat is the example I know best, because I built on it. It puts every major model behind one interface — mid-conversation model switching, document chat with retrieval, image generation, conversation search. It is genuinely excellent software. It also arrives as a repository you clone and a stack you orchestrate: an API server, MongoDB, a vector database, a search index, a retrieval service, a reverse proxy. Six or seven containers depending on what you enable, dozens of environment variables, TLS to terminate — and then API credentials from each provider on top of all that.

If you run infrastructure for a living, that is an afternoon’s work. If you don’t, it is a wall.

All The AIs existed to take that wall down. Not by simplifying the software, which was already right, but by running it for people. You would arrive, start typing, and use good models straight away: no repository, no containers, no provider accounts, no keys, no token arithmetic. Try it free on cheap models to find out whether it suited how you work, then pay one operator directly instead of opening billing relationships with four vendors.

What it was aiming at

A service, not a demo — and that distinction shaped everything about how it was built.

A demo has to impress you once. A service has to be there on a Tuesday afternoon when you are mid-task and depending on it. That means uptime, conversation history that persists, billing that reconciles, and somebody accountable when it breaks.

That is the part it never reached.

The software itself ran, and ran well: 109 distinct models served real traffic across 29 months, through 1,641 conversations and close to 20,000 messages, including a guest mode that let anyone try the whole thing with no signup at all. As a working deployment, it did what it was meant to.

But a working deployment is not a service people can rely on, and the distance between them is not an engineering problem. It is support, on-call, billing infrastructure, trust — a business standing behind it. One person building alongside other work can reach the first. The second needs a company.

Why it stopped

I wound it down through 2025 as other work took over, and retired it properly in August 2026 — data archived, shutdown documented, nothing left dangling.

The gap it aimed at was real, and it has narrowed since — LibreChat now runs a hosted demo of its own, which takes care of the setup problem for anyone who just wants to try a multi-model interface. But a demo is still not a service you would build a working habit on, so the original idea is not dead. It just needs a company behind it, and that was not the shape of what I was doing.

The stack

Customised LibreChat fork. Multi-provider routing via OpenRouter; retrieval-augmented document chat on pgvector; Meilisearch for conversation search; MongoDB; Docker Compose behind nginx. Two independent deployments from one codebase — the main service and a separate anonymous guest instance — with usage metering and per-user balances.