I built Intexere because the platforms I tried were either too limited for what I wanted or built mainly for coding. I also wanted to make my own decisions about how it works. I prefer SwarmUI’s interface to ComfyUI’s, for example, so Intexere has its own image interface and still lets you open the SwarmUI and ComfyUI backends directly when you want them.
I also didn’t want models reaching into different systems on my computer, so
each service runs in its own Docker container. Coding models get their own
sandboxes, where they can run rm -rf as often as they like without breaking
the machine.
And I wanted a challenge. I’m relatively new to software development, and this was a project I thought would hold my interest long enough to stick with.
Intexere is a self-hosted AI workspace that runs on your own computer. It has a FastAPI backend, a React and TypeScript interface, and SQLite for persistent state. Docker runs the application and can manage supporting services such as Ollama, SwarmUI and ACE-Step.
It is under active development, but I use it for everyday work. Chat, coding workspaces, tabletop campaigns and creative tools are already part of the application.
Chat and coding workspaces
Chat supports local models through Ollama or vLLM, hosted models through OpenRouter or Anthropic, and other OpenAI-compatible services. You can attach documents, retrieve passages from them, search the web and change models during a conversation.
For coding work, a model can use a sandboxed workspace to inspect files and run commands. With the appropriate permissions, it can clone a repository, push a branch and open a pull request on GitHub. You don’t need to configure any of this to have an ordinary conversation with a local model.
Long conversations have proved harder than short chats. Loading thousands of messages, choosing useful context and condensing older turns without losing important details have led to changes in how Intexere handles history. Manual compaction still has defects I’m working through, but the application can already continue conversations across model changes and sessions.
Table, Studio and the Library
The Table is for tabletop campaigns. A language model can act as game master, while Intexere records the characters, party and campaign events. Generated narration cannot quietly rewrite an established character or an earlier decision. You can return to an existing campaign and carry on where it stopped.
Studio uses services such as SwarmUI and ComfyUI to make images and video. Music generation runs through ACE-Step, with song titles and cover art alongside the output. The Library holds generated material in one place, where chats and campaigns can use it too.
Intexere can connect to services that are already running or manage supported services through Docker. The managed option is convenient, but I’ve hit cases where the application loses track of an ACE-Step installation and needs to be reconnected. Service management is one of the areas being tested and repaired.
Preparing the public release
The code is MIT-licensed, and a public source repository is being prepared. The current work includes better service management, more predictable context handling and diagnostics that make failures easier to investigate.
What it gives me today is that I can move between a conversation, a coding task and media generation without rebuilding the setup for each one. The next changes are about making that hold up when something fails halfway through a task.