OpenBot

• Giving a Codex agent and a Claude Code agent separate roles on the same project, with one implementing a change and the other reviewing it
• Setting up a research agent with a routine that summarizes what changed in a repository or inbox every morning
• Queuing several tasks, such as drafting release notes and checking new sign-ups, for an agent that is still busy with earlier work
• Having an agent sign in to a web dashboard in the built-in browser and pull data from it
• Running OpenBot on a home server or Mac mini and checking on agents from an iPhone or Android phone
• Moving agents built in Grok Bot to your own computer with their instructions, skills, routines, and memories

OpenBot's main strength is that it treats agents from different vendors as interchangeable workers: a Codex agent and a Claude Code agent can share a channel, hand tasks to each other, and keep their history if you move them to another provider, all on hardware you control and paid through plans you already have. Queues, routines, a built-in browser, and mobile access make it a credible local substitute for hosted agent workspaces. Two things should shape the decision. Agents run with full access and auto-approval by default, and the optional Workspace only mode is not presented as a security boundary, so treat it as you would any unsandboxed automation on your machine. And the noncommercial license means businesses need to talk to the developer before using it for work. For personal projects and experimentation it is easy to try; for company use, settle the license and the security model first.

About OpenBot

OpenBot does not ship a model. It is a desktop workspace that drives the agent tools you already have: the local Codex App Server, Claude Code, and, through the Agent Client Protocol, Grok CLI, OpenCode, Gemini (via Google's Antigravity ACP server), Cursor CLI, and Cline CLI, plus custom OpenAI-compatible endpoints and local model servers such as Ollama and LM Studio. You describe an agent in one prompt, review the generated instructions, and save it. From then on the agent has a name, its own folder under `~/OpenBot/Agents`, a persistent conversation, and a FIFO message queue you can pause, resume, or cancel, so you can keep sending work while it is busy. Agents share files through `~/OpenBot/Shared`, send each other messages and managed file transfers, and work together in channels where one agent owns a task and can explicitly delegate parts of it, with Stop, Resume, and Reassign controls for you. Routines run an instruction on a schedule in the agent's own thread, a persistent embedded browser lets agents open and control pages, and optional Computer Use extends that to the desktop. Switching an agent from Codex to Claude Code keeps its workspace and conversation. A Marketplace offers shared agents, skills, and plugins for services such as GitHub, Notion, Linear, Figma, and 1Password. Workspaces, chats, and browser data stay in a local SQLite database. An account is needed only to invite coworkers to a team host or to reach your agents from the iPhone and Android apps, which connect to your machine over an encrypted peer connection. You can also run OpenBot as an always-on Linux server or Docker container, or rent a hosted server in the EU from €20 or $25 a month. The tradeoffs are significant. OpenBot is a development preview in which agents run by default with full access and auto-approval (a per-agent Workspace only mode limits where they can write, but the developer says full local access is not a security boundary), the code is under the PolyForm Noncommercial 1.0.0 license rather than an open-source one, so commercial use needs a separate license, and the project ships releases almost daily.