open-webui v0.13.0

Open WebUI is a self-hosted web interface for chatting with local or hosted AI models, and version 0.13.0 moves it from a chat window toward a workspace where a model does hands-on work. The main addition is a tabbed panel beside the conversation that can hold files, several terminal sessions, generated documents, a web browser and the desktop of a connected computer. Most of the hands-on features depend on Open Terminal, a companion project that gives agents an environment to run commands, manage files and execute code, either in a Docker container or directly on a machine with the user's own permissions. For anyone who administers an installation, the release also carries upgrade requirements and changed defaults that matter more than any single feature.

The new tabs let a person and a model work on the same machine at once. With a compatible Open Terminal connection that has desktop support enabled, a Desktop tab shows the connected computer's screen and accepts mouse and keyboard input, so the user can follow a model's computer-use work there and step in to operate a graphical program or handle a step that needs manual input. A Browser tab does the same for websites and local development servers: in Chrome mode the user and the model share one browsing session, which lets the model fill in forms and test the web apps it builds. A Git panel can review diffs, commit, manage branches and handle GitHub pull requests when the terminal's folder is a repository. Other additions point the same way: models with native function calling can now create private Python tools in a user's workspace on request, and a model can delegate work to other configured models rather than only to copies of itself.

The operational notes are the part an administrator should read first. The project says the release contains security fixes it has deliberately left out of the changelog until advisories are published after CVE identifiers are assigned, and strongly recommends updating every deployment as soon as possible. It also changes the database schema, so the notes advise a backup first and warn that every instance in a multi-server or load-balanced setup must be updated simultaneously, because rolling updates will fail. Two defaults loosen limits on agent work: a single response can now run up to 1,024 rounds of tool calls instead of 256, and sub-agent limits that previously defaulted to 20 concurrent, 20 in the background, 30 tool loops and 30,000 characters of output now default to no limit. Both can be set back through environment variables, so administrators who want the old caps now have to set them explicitly.

The release notes describe intended behavior and do not show how these features perform in practice. The security fixes themselves cannot be assessed until the advisories appear. This explanation is based on the 0.13.0 release notes and the Open Terminal repository description; the linked commits and the separate Desktop app changelog were not read.