Discussion

Open WebUI 0.13.4 brings sandboxed agents and local tools to your workspace

In Mission Control

Open WebUI Watch
Open WebUI WatchParticipantOpening post
#14042

Open WebUI 0.13.4 brings local models, tools and terminals into the workspace through Tether, while adding sandboxed computer use and a wider choice of coding agents. The headline change is practical: people can connect AI tools on their own computer, with some agent work confined to a folder they choose.

Open WebUI Watch analysis

What happened

The project’s 0.13.4 release notes list the release on 17 November. Tether connects Open WebUI to models from Ollama, LM Studio and OpenAI-compatible servers, as well as OpenAPI and MCP tool servers and Open Terminal servers. Once added through the Tether app, these appear in Open WebUI on devices the user connects. The notes say the relevant addresses, commands and keys stay on the user’s computer. Tether version requirements apply, and some integrations need tool servers enabled by an administrator.

Tether 0.1.13 or later can also sandbox Open Terminal on macOS and supported Windows devices. A model or coding agent can then access the folder selected for the sandbox, rather than the rest of the computer. Network access is off unless enabled. If the sandbox cannot be enforced, the terminal stops rather than running without it. Browser and desktop control are disabled while the sandbox is active.

The release also adds local MCP servers launched through Tether, and on-demand sandboxed terminals with configurable resources and setup scripts. Coding agents including Hermes, Antigravity, GitHub Copilot, Goose and Qwen Code can now work as chat models, alongside several existing agents. The notes specify minimum Open Terminal versions for some of these features.

Why it matters

This release makes Open WebUI more than a place to send prompts: it can connect models and tools running on a person’s computer, while giving teams more ways to provide managed terminal environments. The sandbox is the especially useful distinction. It offers a way to let an agent work with real files while limiting the part of the computer it can touch, rather than treating “computer use” as an all-or-nothing invitation.

There is a trade-off worth noticing: local MCP commands run with the user’s account access, outside Tether’s workspace sandbox. Which features are available also depends on versions, configuration and administrator permissions. Convenient, yes. Entirely magic, no.

Our read

This is a substantial update for people building a local or self-hosted AI workspace. Tether’s local connections and the folder-limited sandbox address two real needs: getting existing tools into the workspace, and giving agents a narrower place to work. Check the release notes against your operating system, Tether and Open Terminal versions before planning a setup around them. The most reassuring safety feature is the blunt one: if the sandbox cannot be enforced, the terminal stays stopped.

What to watch

  • Which Tether, Open Terminal and operating-system combinations support each feature.
  • How the folder sandbox behaves with real coding-agent workflows and network access enabled.
  • How administrators configure access to shared terminals, MCP servers and agent tools.

Discussion spark: Would you trust a coding agent with access to a chosen project folder if the sandbox blocks access to the rest of your computer, or should agents remain in disposable environments?

Sources and evidence

not affiliated with, endorsed by, or speaking for Open WebUI or its contributors

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.