Intel / repo

open-webui v0.12.4

Source: github.com
Author: github-actions[bot]
Date: Published Oct 25, 2026

From the Daily Brief

Open WebUI 0.12.4 adds an admin setting, ENABLE_SKILL_LAZY_LOADING, that lists selected skills for the model and sends their full instructions only when the model asks to load one. Skills mentioned directly in a message still load right away. The option is off by default and needs the built-in Skills tool. The release also makes context compaction visible: a Context compacted line marks where older messages were summarized and opens to show the summary, including when compaction ran between tool calls inside one reply. Compaction now runs in temporary chats and in short chats that are already over the threshold, such as a single question whose reply is full of tool output. Before this, the follow-up, regenerate or edit in those chats failed with a context length error. Administrators can also hide individual controls from users.

Read the 2026-10-26 Brief →

Context

Version 0.12.4 of Open WebUI, a chat interface for language models that administrators install on their own servers, for example from its Docker image, contains security fixes that the project says are not listed in the changelog (GitHub release notes). It strongly recommends updating all deployments as soon as possible and says it will publish the advisories once CVE identifiers are assigned. Administrators therefore know the update is urgent before they can see what it fixes.

The same update changes behavior that can break existing setups. Direct Integrations, a setting that is off by default, used to only hide a settings tab when off; now, while it is off, the tool servers and Open Terminal connections that users and admins saved are no longer used in chats. Instances that never turned it on lose those connections until an admin enables it in Admin Settings, Connections, or sets ENABLE_DIRECT_INTEGRATIONS. The Docker image also moves from Python 3.11 on Debian 12 to Python 3.12 on Debian 13, so custom functions, tools and pip packages installed inside the container need versions that support 3.12. Sharing permissions now apply to sharing through the API as well, and a disallowed change to an item's access is refused with an error instead of being quietly dropped.

The headline feature addresses how much text each chat sends to the model. Skills are reusable instruction sets that can be selected in a chat or attached to a model, and normally their full instructions travel with every message. With the new on-demand option, which is off by default and needs the built-in Skills tool, selected skills are only listed, and the model loads one when it needs it; skills mentioned in the message still load right away. A related change runs , which summarizes older messages when a conversation gets too long, in temporary chats and in short chats already over the limit, such as one question followed by a reply full of tool output, so follow-ups there no longer fail with a error.

Key takeaways · AI-distilled

Open WebUI v0.12.4 includes security fixes it does not list in the changelog. The project strongly recommends updating all deployments as soon as possible and says advisories will be published after CVE assignment.
Upgrade check: while Direct Integrations is off, which is the default, tool server and Open Terminal connections saved in users' and admins' Integrations settings are no longer used in chats until an admin turns the setting on.
The Docker image moves from Python 3.11 on Debian 12 to Python 3.12 on Debian 13, so custom functions, tools and pip packages installed in the container need versions that support 3.12.
Tools from tool servers, including Open Terminal, now reach the model in the same form after every restart, so the provider's prompt cache keeps working and long chats are no longer reprocessed from scratch.
_INITIALIZE_TIMEOUT, 10 seconds by default, now also limits tool listing and resource reads, so slow MCP servers may need a higher value; MCP connections also add support for the stateless 2026-07-28 protocol.

Terms in this piece · Glossary

context compaction — Summarizing an agent's earlier conversation to free room in the context window so a long session can keep going.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.

Read the source github.com