frontier-ai-next/umbrella

Umbrella is a control panel for running AI coding agents through structured, multi-step tasks called workspaces. Rather than letting one AI agent run freely from start to finish, Umbrella breaks a task into phases, such as research, planning, execution, review, and verification, and for each phase it starts a separate agent to do the work along with a second agent whose only job is to watch for the first one stalling, repeating errors, or going over budget. A human can check in on the run through a chat interface at any point, or restart it with feedback after it finishes. The project can run different underlying agent engines interchangeably, which it calls heads. It ships with its own engine called Ouroboros built in by default, and supports plugging in an external one called Hermes as well, and it can even run several heads against the same phase at once and pick whichever one produces the best result based on the evidence each one collected. Memory is a major focus: the system keeps track of what has worked before, both within a single workspace and across different workspaces, using one of three interchangeable storage backends behind a shared interface. Each phase in a run follows a written manifest file that spells out exactly which tools, skills, and instructions that phase's agent is allowed to use, so the system does not let an agent do things outside its current step. The project includes both a command line way to start a single run and a web interface, complete with its own local server and a React based dashboard, for watching runs, browsing memory, and adjusting settings. It can also be deployed with Docker for hosting outside a local machine. Running it requires Python 3.11 or newer, the uv dependency manager, and Node.js with Yarn or npm to build the web interface, plus your own API key for whichever language model you connect it to.