← All projects

Self-hosted big-AGI

My own multi-model AI chat, on my own domain

Live 2025 - Present Fork maintainer and operator

A self-hosted fork of big-AGI running on Vercel under my own subdomain, wired to several model providers. It replaced a browser extension I used for querying multiple models side by side, and it is where a lot of my day-to-day model comparison happens.

Highlights

  • Deployed on Vercel with provider keys kept server-side, so the browser never holds a credential.
  • Multi-provider setup through OpenRouter, which makes comparing models a configuration change rather than a new subscription.
  • The beam feature (querying several models at once and merging the answers) is what made me drop the extension I was using before.

Built with

Next.js TypeScript Vercel OpenRouter

What it is#

A self-hosted fork of big-AGI, deployed on Vercel under a subdomain of this site, wired up to several model providers through one interface.

It is not a product I built from scratch. It is infrastructure I run, and the reason it is on this list is that operating something is a different skill from writing it.

Why self-host it#

I was using a browser extension to query several models side by side and compare answers. It worked until it did not: extensions break, get abandoned, and put your conversations somewhere you do not control.

Self-hosting solved three things at once:

  • One interface, many providers. Routing through OpenRouter means adding a model is a configuration change rather than another subscription. That matters when you are comparing models seriously rather than picking one on faith.
  • Keys stay server-side. The browser never holds a provider credential. This is the part that most casually-deployed AI frontends get wrong, and it is the reason I would not have used someone else’s hosted instance.
  • The beam feature. Querying several models at once and merging the responses is the thing that replaced my extension outright. Being able to see three models disagree about the same question is far more informative than trusting one.

What running it actually taught me#

Deployment is a real part of the work. Vercel makes this easy, and easy is not the same as free of decisions: environment variables, which routes execute on the server, what gets cached, and what happens when a provider returns an error mid-stream.

Forks age. Upstream moves, and a fork that is never updated slowly becomes an unmaintained copy of someone else’s old code. Deciding when to pull changes in, and what to do when local modifications conflict, is ongoing work with no completion date.

Model comparison is a habit, not an opinion. Having this running changed how I choose tools, because trying something became cheap. Most of what I now believe about which model to use for which task came out of this interface rather than out of benchmarks.

How it is built#

Next.js and TypeScript (upstream’s stack), deployed on Vercel, with model access through OpenRouter and direct provider keys.

This is also, honestly, most of my hands-on Next.js and Vercel exposure. I did not write the application, I run it and modify it. I would rather say that plainly than let a stack list imply more than it should.

Status#

Live at agi.reizen-desu.com, and part of my daily workflow.