Our team does most of its work with Codex and Claude Code in Open WebUI: software engineering, documents, and everything in between. The agents run on a machine, not on the device in your hand, so we can start a job at a desk and follow it, approve it, or hand it to a teammate from a phone or tablet.
You can run Codex or Claude Code in Open WebUI and use it from any device you're signed in on, including your phone. That's how our team does most of its engineering and writing now.
Until recently, using an agent like Claude Code or Codex meant sitting at the one computer it was installed on. If you stepped away, the work stopped with you. Nobody else could pick it up, and you couldn't check on it from anywhere else.
In Open WebUI, the agent runs on a computer connected to Open WebUI, not on the device in your hand. Each chat keeps its machine, folder, and open files on the server. A job you start at your desk opens the same way on your phone, where you can follow it, approve its next step, or tell it what to do next. A teammate can take over a chat halfway through. And the same agents handle documents as readily as code.
This is how that works, for one person, for a team, and for an organization. The clips show a demo team working on example projects.
Software engineering #
A lot of engineering work starts from something already written down: an issue, a failing check, a review comment. Instead of copying it into a prompt, you can hand it to the agent from where it already is. With the GitHub CLI signed in on the machine, the Git panel in the Files tab lists the repository's issues and pull requests. Work on this on an issue writes the request for the agent, with the issue, its comments, and instructions to branch, make the change, and summarize it. On a pull request, Fix hands over a failing check, and Address feedback hands over the review comments. You read the request, pick Codex or Claude Code, and send it.
While the agent works, the chat shows each command it ran with its output, and a diff of every file it edited. When it's done, you review the changes in the same panel, then commit and push or open a pull request, without leaving the chat.
Run gh auth login once in the machine's terminal to set it up. Issues need Open Terminal 0.17.8 or later.
Writing and documents #
The same agents that fix code also write and fix documents: docs, READMEs, release notes, reports, spreadsheets. They work on the files themselves, so the result is a finished file, not text you copy across. Spreadsheets and Markdown open in an editor in the Files pane, and they save straight back to the machine.
The changed lines in a file beside the chat, then a spreadsheet cell fixed and saved back to the file
When only one passage needs to change, select it in the open file, and an Ask or edit box appears next to the selection. The request goes to the chat with that passage quoted and the file attached, so you never have to describe where the text is.
Maya selects an out-of-date paragraph in the team runbook and asks for a rewrite in Ask or edit, and the model reads the script it describes and rewrites just that paragraph
From your phone or tablet #
On a phone, most of the work is following along and answering. With Tool Permissions on and the chat set to Ask, an action that needs approval waits in the conversation, so you can approve a push from wherever you are and read the result later. If your connection drops, the agent keeps working for up to an hour. A message you send while a reply is still running waits in a queue. In your own saved chats, that queue is kept on the server, so it survives closing the page or switching devices.
A request started at the desk waits on the phone, and approving it there lets Codex commit and push
Hand a job to a teammate #
A half-finished job doesn't have to wait for the person who started it. Share the chat with Allow replies, and your teammate sees the whole session: every message, every command and its output, and every approval. When they reply, the agent continues its own session, on the same machine and files, with everything it already knows. Someone reviewing the change can ask the agent why it did something, and the answer comes from the session that did the work.
For a whole project, share the folder. Its page can show the chats as a board, with columns by status, so everyone can see which agents are done and which are waiting on someone.
Maya shares her Codex chat with replies allowed, and Sam reads every step the agent took and has it finish the job on the same files
Work that runs on its own #
Some jobs only matter if they happen regularly, such as a nightly check of a project's dependencies. An automation runs an agent on a machine and in a folder, on a schedule, and posts each run as a new chat in the folder. In a shared folder, everyone who can edit it can pause the automation, run it, and read its runs.
Administrators can always create automations. Everyone else needs the automations permission, which is off by default. Agent automations post to a chat in the folder, not to a channel.
Maya has Codex check the project's dependencies every night, clicks Run now, and opens the run's chat and the report it wrote
Where the agents run #
Where an agent runs decides who can pick its work up. On a machine the team shares, with the agents signed in once, anyone with access can carry on a job. Terminals gives each person, chat, or model a machine of its own, made from the chat, inside limits an administrator sets. Tether, our desktop app, connects your own laptop. The agents installed on it then appear wherever you sign in, working on the laptop's own files, as long as the laptop is on.
A project folder in the sidebar can hold a default terminal, a working folder, and an agent, so every new chat in it starts in the right place. On a small screen, that means you can ask straight away, without picking three things first.
An agent on your laptop works on your real files, so decide first which of its actions need approval. How to let your model run commands and edit files in Open WebUI sets up each kind of machine, and How to run agents like Claude Code in Open WebUI covers the agents and their approvals.
What to hand off first #
Pick a project folder with its machine and agent set, and start with something you'd otherwise do yourself:
Work on issue 128: make the change on a new branch, run the tests, and summarize what you did.
Rewrite the installation section of the README so a first-time user can follow it.
Every night, check whether any of this project's dependencies have new versions, and list what changed.
Total this spreadsheet by month and add a chart on a new sheet.
Everything here works on Open WebUI v0.13.4. If you've moved some of your own work into Open WebUI, we'd like to hear how on Discord.
The Open WebUI Team
AI on your terms.
Your models. Your data. Your infrastructure.




