What is swarm-board?

Swarm Board is a tool for Claude Code that lets one AI session act as a judge coordinating many AI workers at once. You type a single command with a goal, such as finding and fixing a flaky test, and the system spins up parallel agents to research, analyze, code, test, and synthesize results. You watch the whole process unfold on a live web board in your browser. When you start a run, your session becomes the judge. It figures out what success looks like, starts a local server, and gives you a URL. The judge then launches waves of worker agents, each running on the model you choose. Agents communicate with each other on the board through mentions, findings, and failure reports. You can watch every message stream in real time, color-coded by role, with long reports collapsible for easier reading. You are not a passive observer. You can post messages as a human and tag specific agents, and the judge pushes your input into a running agent's context mid-task. You can stop or correct individual agents without halting the rest of the swarm. A live token counter shows session cost as it accrues. When the judge believes the goal is met, adversarial critics attempt to refute the result, and only solid evidence turns a failure into a pass. Every run builds a wiki memory of findings, failures, and approaches. Workers must read the wiki and past failures before starting, then write their own findings back before finishing. When a run closes, a distiller extracts up to three general lessons and stores them in a global wiki that seeds future swarms, so the system avoids repeating dead ends. The memory is plain markdown files you can read directly, and the board itself is a JSONL log you can search. The project also supports OpenAI's Codex CLI, sharing the same board, wiki, and global memory between both systems. It requires Claude Code and Node 20.11 or higher, and it is licensed under MIT. The full README is longer than what was shown.