OpenBot explained: run Codex, Claude Code, and Grok in one app

OpenBot has been climbing GitHub Trending fast. It is a desktop application that collects Codex, Claude Code, Grok, and other command-line AI tools into a single window, gives each one its own working directory, queues tasks, and lets them hand work to each other.

The repository was created in August 2026. In one week in late October its stars went from 690 to more than 2,400, and on October 28 alone it gained 1,053.

Plenty of people now hold more than one AI coding subscription: Codex through ChatGPT, Claude Code through Claude, Grok CLI through xAI. Running several at once means several terminal windows that know nothing about each other, and closing a window throws away its history. If you want agent A to hand its findings to agent B, you copy and paste.

OpenBot wraps those command-line tools into one application. It brings no model of its own; it uses the CLIs you already have logged in locally, then gives each agent a name, an avatar, instructions, a working directory, and a conversation record. Close the app and reopen it, and all of that is still there.

Codex and Claude Code are supported natively. Grok CLI, OpenCode, Gemini, Cursor CLI, and Cline CLI connect through the ACP protocol. If Ollama or LM Studio is running locally, OpenBot detects them at startup, so local models work too.

Models are set agent by agent: which provider to use, how far to open the reasoning dial, all up to you. You can switch providers midstream as well, say from Codex to Claude Code, and the workspace and the previous conversations carry over. Nothing starts from scratch.

Files live in ~/OpenBot/Agents/<agent-id>. When agents need to share files, they go through ~/OpenBot/Shared, and what passes through carries a manifest and checksums. Conversations and the task queue sit in a local SQLite database.

So agents accumulate. What they chatted about, which files they changed, all of it is still there when you reopen the app and pick up where you left off.

Messages are handled in order, first in, first out, and you can pause, resume, or cancel them. If the app crashes and you reopen it, the queue survives. Send a new message while an agent is still working on the previous one and it simply waits its turn.

Ordinary terminal windows cannot do this. Two agents can message each other directly, react with emoji, and transfer files. One agent finishes looking something up and sends the result straight to the agent that writes the code, with no human relaying in the middle.

A channel is a common chat area for several agents. One task has exactly one owner at a time, you can assign work to a specific agent explicitly, and the interface carries Stop, Resume, Reassign, and Archive buttons. Splitting a job across agents, research to one, drafting to another, proofreading to a third, is exactly what a channel is for.

The app ships with a browser of its own. Agents can open pages in it, read the content, and operate the page. There is also an optional Computer Use feature on macOS, Windows, and Linux; the driver comes with the installer, so there is nothing extra to install. On macOS you need to grant Screen Recording and Accessibility permissions. Beyond that, the set of applications it can connect in one click is fairly broad.

When a conversation runs long and approaches the model's context limit, the app compresses it automatically; there is no manual cleanup. Agents can also save workflows they use often as skills in a local skill library, and other agents can install and use those skills too.

There are clients for iPhone and Android. They connect to the OpenBot instance on your own computer, which means the computer has to stay on. If you want it always available, install it on an always-on server instead: the project provides a one-click script and a Docker image, and you then connect from the phone or from a web browser.

Three things, before anything else.

1. The permissions are wide open. The project is still a development preview, and agents run in danger-full-access mode. After you agree once at first launch, reading and writing files, running commands, going online, and operating the browser never ask again. Our suggestion: run it on a spare machine, in a virtual machine, or in Docker for a while before putting it anywhere near your main computer.

2. Check the license. It is PolyForm Noncommercial 1.0.0. Personal and non-commercial use is free; commercial use requires a separate license. Versions before 0.1.11 were Apache-2.0. If a company wants to put this to work, settle this question first.

3. Network. Local-first means your data is stored on your machine, but as long as the agents use cloud models, the conversation content still goes to OpenAI, Anthropic, xAI, and the like. To stay off the network entirely, connect local models through Ollama.

If you already subscribe to one or two of ChatGPT, Claude, Gemini, or Grok and regularly keep several agents open at once, installing this costs you little; OpenCode even offers free models without login. If you need a cloud setup that runs with the computer off, or your company has hard requirements around audit and sandboxing, watch from the sidelines a while longer.

One more thing worth knowing, since OpenBot drives CLIs rather than models: any CLI session configured against the DeepSeek API becomes a DeepSeek-powered agent inside the same window. Claude Code and OpenCode both take a DeepSeek endpoint, and this site documents both setups. One subscription can sit behind every DeepSeek-backed session you run in there.

Several agents running at once spend tokens faster than one. DeepSeek Coding Plans pair an official key with a direct API connection at a discount, in one-off and explicitly selected recurring options, while stock lasts.