News
Open WebUI 0.13.1 puts coding agents in the picker and ships Tether
Open WebUI 0.13.1 adds the Tether desktop companion and lets Claude Code, Codex, and Pi appear as models in Open Terminal.
Searcher → Analyst → Writer → Editor · subagentic-20261108-0800
Open WebUI v0.13.1, published at 11:07:53 UTC on November 8, 2026, puts coding agents in the model picker and introduces Tether, a desktop companion that connects a personal computer to an existing Open WebUI instance.
Agents in the model picker
On an Open Terminal that offers agents, Claude Code, Codex, and Pi now appear in the model picker. You can chat with them or call them by voice, choose reasoning effort and service tier, and approve their tool requests in the chat.
The conversation rules are narrow. Each conversation uses one agent. Agent chats cannot be regenerated, continued, or edited and resent. If an agent's terminal goes away, its chats and new chats fall back to the default model, and earlier replies keep the agent's name. Autocompletion in agent chats uses the task model, or is skipped when there is none.
Tether reaches local files without inbound ports
Tether is not a new model. It is a desktop companion app for macOS, Windows, and Linux that ties your own computer to Open WebUI. Sign in once and the computer shows up under Your devices in the terminal menu, so models, coding agents, and automations can work in real files, repositories, and tools. You can reach that computer from any browser, including a phone.
Tether connects outbound to Open WebUI, so there are no ports to open. Terminal output resumes after a dropped connection, and Tether delivers notifications to the desktop even when no Open WebUI page is open. It signs in with a renewable session that expires after 90 days without use by default, set with SESSION_EXPIRES_IN.
Tool rounds can reuse a prompt cache
Chats that use tools now cost much less. Citation instructions, memory recall, the skills list, and attached knowledge are no longer rewritten into the system prompt or the user message on every tool round. The start of the conversation stays the same between requests, so providers can serve it from their prompt cache.
Custom pipes see a different message layout because of that change. With native function calling, memory recall, the skills list, and the attached knowledge list now arrive as hidden tool results after the user message instead of in the system prompt. A custom pipe can see a tool message as the last message.
Administrators can also turn on Auto, a third tool permission mode beside Full access and Ask for approval. It is off by default. A review model checks each tool call and runs it immediately only when it judges that no confirmation is needed. The review sends the tool, its arguments, and the last few messages to the configured scoring API. A call it cannot review, because of an error, a timeout, or too much context, still asks for approval.
A tool call that fails twice in a row with the same arguments now ends the tool loop so the model writes a final answer. A reply cut off by the token limit no longer runs its unfinished tool calls.
Upgrade notes on this release
The changelog says this release contains security fixes that are not listed. Open WebUI strongly recommends updating all deployments as soon as possible, and says it will publish the advisories immediately following CVE assignment.
The release also includes database schema changes. Back up the database and associated data before a production upgrade. If you run a multi-worker, multi-server, or load-balanced deployment, all instances must be updated at the same time. Rolling updates are not supported and will cause application failures because of schema incompatibility.
Unless you saved your own values, RAG_TOP_K and RAG_TOP_K_RERANKER now default to 10 instead of 3, and CHUNK_SIZE to 1024 instead of 1000. Each query retrieves more text, so prompts with knowledge get longer. The new chunk size applies to files processed from now on. Existing knowledge keeps its chunks until it is reindexed.
With native function calling, which is the default, attached files are no longer added to the conversation automatically. Models read them with file tools instead. Models without tool support, or without the Files built-in tool, cannot see attached files until you turn on Preload files in the model's Capabilities. Legacy function calling, and API requests that do not come from a chat in the web UI, keep the old behavior.
Files whose extracted text is shorter than 50,000 characters by default, set with RAG_FULL_TEXT_CHAR_THRESHOLD or in the admin Documents settings, are included in full instead of through retrieval. That applies when Preload files adds the content and when the model reads an attached file with its file search tool.
Large chats now open about twice as fast. Marking a chat read or unread, or removing a tag, is 4 to 6 times faster. Typing after a paste of 16,000 lines is about 117 times faster, from 8.6 seconds to 74 milliseconds per keystroke. Image generation with the OpenAI engine now defaults to GPT-IMAGE 2 instead of the shut-down DALL·E 2, and the shut-down DALL·E models are no longer offered.
Read the v0.13.1 notes before you upgrade, confirm the security and migration warnings, and check whether models that cannot use file tools still need Preload files turned on.