Open WebUI can now run coding agents like Claude Code and Codex from its model picker, with tool-request approvals in chat. Tether links your own computer so agents work in your real repositories.
Key takeaways · AI-distilled
Tool chats cost less because citation instructions, memory recall, the skills list and attached knowledge are no longer rewritten into the system promptThe standing instructions a model receives before any user input — defining its role, rules, tools, and tone for the whole conversation.Full definition → each tool round. They arrive as hidden tool results, so the prompt prefix stays stable and providers can cache it.
A new opt-in Auto tool permission mode has a review model check each tool call and run it only when it judges no confirmation is needed. It sends the tool, its arguments and your last few messages to the scoring API, and any call it cannot review asks for approval.
With native tool useA model's ability to call external functions — run code, search the web, edit files — instead of only generating text.Full definition →, attached files are no longer preloaded into the conversation. Models read them with file tools instead, so a model without tool support or the Files built-in tool cannot see attachments until Preload files is turned on.
Retrieval defaults changed unless overridden: RAGRetrieval-augmented generation — fetching relevant documents first and pasting them into the model's context so it answers from your data instead of memory.Full definition →_TOP_K and RAG_TOP_K_RERANKER go from 3 to 10 and CHUNK_SIZE from 1000 to 1024, so knowledge prompts get longer. Files under 50,000 extracted characters are now included in full instead of through retrieval.
The release ships unlisted security fixes and database schema changes. The maintainers recommend updating soon, backing up first, and updating every instance of a multi-worker or load-balanced deployment at once, since rolling updates will fail.
Terms in this piece · Glossary
system prompt — The standing instructions a model receives before any user input — defining its role, rules, tools, and tone for the whole conversation.
tool use — A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.
RAG — Retrieval-augmented generation — fetching relevant documents first and pasting them into the model's context so it answers from your data instead of memory.
