CerebrumKit is a self-hosted starting point for agentic projects: build agents out of skills and tools in an admin panel, give them your own database tables, and let a start / group / stop workflow decide which agent answers. Tools and skills are database rows - edit one and the next message uses it, with no restart and no deploy. FastAPI + Vue + Postgres or SQLite, Apache-2.0.
Hi Product Hunt. I build AI agents for businesses, and the first month of every project was the same month: an admin panel, a tool registry, somewhere to put the client's data, a way for the client to edit prompts themselves, and a login for their staff. None of it was the interesting part.
CerebrumKit is that month, done once and open source.
The idea that makes it different: everything about an agent is data. A tool is an OpenAI function spec plus a Python body, both stored in the database. A skill is written instructions plus the tools it needs. An agent is a description that is its system prompt, plus its skills. Editing any of it in the admin panel takes effect on the next message - no restart, no deploy. The people who own the process can change it without me.
Around that: a storage library of your own tables with Excel import/export, chats over websockets, memory per (agent, user), delegation between agents, an admin/client role split so the client sees only their own projects, and six UI languages including the documentation.
Two things I want to be straight about, because they are the honest limits: tool bodies are Python executed on your server and are not sandboxed (so tool authoring is admin-only), and it is early - no test suite yet, no hosted demo. It runs on Python 3.12+, Node 20+ and Postgres or SQLite.
Free, self-hosted, and I will be here all day for questions - especially the hard ones about where this design goes wrong.