I was running Claude Code, Codex and a couple of other CLI agents in a pile of terminal tabs, and kept losing track of which one was doing what. I wanted one place to give them work, see what they were doing, and step in only when something actually needed me like a small team with a Chief of Staff, not a wall of terminals. The problem I kept running into with other "AI agent" tools was that they wanted me to hand my API keys and logins to their server. So I built OpenBot the other way: no hosted backend, no bundled credentials. You sign in to each engine (Claude Code, Codex, Cursor, Gemini CLI, Grok Build, or a fully local model via Ollama/LM Studio) with the CLI you already use, and everything stays on your machine. The approach evolved a lot around trust: early versions just ran whatever a Bot asked for, which felt unsafe. I ended up building a fast decision layer (Jev) that checks every risky step and asks for approval in plain language the exact command and a risk level, not a wall of JSON before anything happens outside the shared workspace. That safety layer became as central to the product as the multi-agent part.

Open Bot

Desktop app for a small team of persistent AI Bots

Desktop app for a small team of persistent AI Bots