BrainSoup

BrainSoup is a Windows desktop platform for assembling custom AI agents that collaborate in chat rooms, use tools, and draw on your own documents, backed by cloud or local language models with local data storage.

BrainSoup is a native Windows application for building a personal team of AI agents, each configured for a specific job. Instead of one general chatbot, users define specialized agents and let them coordinate inside goal-oriented chat rooms. Routine tasks and larger projects are handled by dividing work between agents with distinct skills and data access, which keeps roles clear and limits the reach of any single agent. Instructions are given in plain language, so the workflow reads more like messaging colleagues than writing configuration code. The platform runs on large language models from providers such as ChatGPT and Mistral AI, and it can also connect to local models through Ollama. Semantic Kernel sits underneath to extend what the base model can do, and the agents keep a sense of time plus memory of the user and earlier conversations. Knowledge boundaries are set by the user: documents can be supplied as reference material, and the agents' knowledge expands as conversations continue. A built-in agent called Doc-E demonstrates the concept and answers questions about setting up BrainSoup itself. Agents are not confined to talking. Built-in tools let them calculate, browse the web, send email, make API calls, generate and edit images and video, and run scripts, with ChatGPT plugins adding further reach. Custom tools can be pointed at programs, commands or scripts on the local machine. Integration with a software environment works through simple text files and scripts in any language rather than complex APIs or SDKs. A sandboxed file system governs how agents and users touch files, and image and audio input are supported, with simpler models borrowing multimodal ability by collaborating with other models. Privacy is the design center: a local database, optional local models, no cloud storage, and third-party requests that are anonymized and not used for training. The audience is broad, covering developers who want scriptable automation, marketers and researchers handling data-heavy work, project managers, students and small teams that prefer a controlled desktop setup over a hosted service. It is an early product, with a command-line interface, additional LLMs, event and context sources, MacOS, Android, iOS, a headless server edition and optional encrypted cloud sync listed as future work.