If v4 was about what Xeref agents can perceive, v5 is about control — over your tools, your models, and the workspace itself. Six releases in under three weeks added a proper tool manager, a wave of new models, real navigation, and the first groundwork for autonomous actions.
v5.5 — AI Task Tools and a Readable Inbox
AI task tools: Tasks now has AI Decompose (break a goal into an ordered, actionable task list) and AI Prioritize (rank your open tasks and suggest what to do next), both available from the toolbar.
Telegram bot settings: Settings → Telegram now has a token-based connect flow — create a bot with @BotFather, paste the token, and Xeref activates the webhook for you.
Readable inbox: Ingested emails now store their full body instead of just a snippet, with a read state so you can open and read a message in place.
Computer: The Tasks tab is now called Computer, laying groundwork for autonomous actions. And Pro's default chat model was bumped from Sonnet 4.6 to Sonnet 5.0.
v5.4 — Back & Forward Navigation
A Back/Forward button pair now lives in the sidebar header, so you can retrace your steps across chat, tasks, stats, and every in-app view. Alt+← / Alt+→ and your mouse back/forward buttons work too. Xeref Desktop also gained a native-style draggable titlebar with window controls, replacing the bare OS frame.
v5.3 — Native Fullscreen
Press Ctrl+F11 anywhere in the dashboard to enter or exit OS-level fullscreen. Bare F11 is left to the browser; Esc exits at any time.
v5.2 — Reliability
Two fixes worth noting: creating an empty artifact from the Artifacts page no longer fails, and email triage jobs now fall back to a safe classification when the model call fails, so the agent fleet no longer accumulates failed runs.
v5.1 — New Models and Direct Anthropic
GPT 5.4 joins Pro and Ultra, and Sakana AI's Fugu Ultra joins the Ultra lineup for long-horizon, high-quality generation.
Antigravity workflow runs: Workflows can now dispatch a step to Google's Antigravity managed agent, running as a background Hermes job so long agent runs never hit a request timeout.
Direct Anthropic (beta): Settings → Xeref Code gained Ultra-only toggles for Server-side Compaction, the Advisor tool, and Cache Diagnostics — routing long Claude coding sessions through Anthropic's first-party API for automatic context compaction, with per-session cache diagnostics for debugging slow turns.
Also fixed: pasting or attaching an image now correctly triggers vision-capable model routing, and Gemini-generated images survive a page refresh — they're uploaded to storage and referenced by URL instead of an inline data URL.
v5.0 — The Server Tools Manager
A dedicated panel (open it via "Manage server tools…" in the + menu) shows all seven server tools — including coming-soon and plan-gated ones — with real on/off defaults that persist across sessions. Web Search, Web Fetch, and Datetime are available now; Image Generation, Advisor, Subagent, and Fusion are previewed as coming soon.
The + menu quick-toggles are now session-only overrides: they affect the current chat and are never saved, while defaults live in the manager. The composer got decluttered along the way — the always-on tool pills and duplicate manage icon are gone.
And Gemini Image generation arrived: select "Gemini Image" in the chat model picker (Pro and Ultra) to generate images from a text prompt.
The through-line of v5 so far: fewer surprises, more levers. Your tools have real defaults, your models match your plan, and the Computer tab hints at where this is heading next.