You can now choose a Google Font for the embedded widget and floating Ask AI button, helping the assistant better match the site it is installed on. The Ask AI button also supports bold and italic text styling.

New API keys now use a cc_live_ prefix, more entropy, and hash-only storage with safe display metadata. Full keys are shown only once when created, while the dashboard stores and lists only non-sensitive prefix and last-four information.

API key authentication now supports both new hashed keys and existing legacy keys, and key deletion is scoped to the owning user to prevent cross-account deletion.

Subpages in the app now have clearer back navigation, making it easier to return from detail and creation screens such as actions and articles. Knowledge tables were also refined so rows are easier to open, actions are easier to find, and long titles behave better.

Questions now appear in the chat widget immediately after submission, so users no longer wait on thread creation or answer startup before seeing their message. The loading state was simplified from skeleton blocks to clear status text like "Thinking..." and "Answering...".

The Web embed page also includes a test widget, so teams can try the real embedded chat experience before installing it on their site.

CrawlChat now resolves source links for answers with a single batched lookup instead of repeatedly querying each search result one at a time. This reduces answering latency, especially for questions that match many source chunks, while preserving citation URLs from the knowledge base.

Added Z.ai GLM 5.2 as a selectable model for collections and channels. The model is available through OpenRouter, is configured for documentation RAG with four retrieved sources, and costs four credits per message.