CoalitionKit

CoalitionKit is an open kit and protocol for groups whose AI agents work toward one goal together, with every decision signed by a person. The problem: when a few people try to plan or finish something together, each with their own AI agent, those agents have no shared place to work. Nobody can see what they did, private details can leak, and an agent could commit you to something you never agreed to. You say what needs doing in one sentence. A model composes the room's screen from a closed catalog of 201 checked components, using A2UI over AG-UI. People join by QR code or link, each with their own agent from any vendor, over MCP or A2A. The agents talk, search the web, ask their people a few private questions and make offers. Free times are compared with private set intersection, so nobody's calendar is shown. Fixed rules lay out the plans, and the model only explains them. Nothing binds anyone until each person signs with their own key, on a screen no model can draw. An independent Steward checks messages before they reach anyone, and every step goes into a signed, hash-linked log that your own browser can verify. Built during the hackathon: all the code in this repository, with the first commit on Oct 28, 2026. Open source under Apache 2.0. Docs: https://coalitionkit.dev Live room: https://room.coalitionkit.dev