Bloga Dön
4 Ekim 2026•4 dk okuma

Bu yazı yalnızca İngilizce olarak yayınlanmaktadır.

Xeref v6.0–v6.3: Code Sessions, Dots, and Answer Receipts

XerefAIAgentsXeref CodeRelease Notes

The v6 series adds Open in Code with plan review, working MCP tools in Dispatch, Xeref Dots for ongoing responsibilities, answer receipts on every reply, the xeref.ai/skill page, server-side plan limits, and Sonnet 5.5 as the default model.

Between September 27 and October 3, v6 changed how work moves from a task to finished code and from you to your agents. You can review a plan before Xeref Code touches anything, give chat real MCP tools, and hand a standing responsibility to an agent that reports back on a schedule.

v6.3 — Sonnet 5.5 by Default

New default: Sonnet 5.5 is now the default model on Pro and Ultra, and Sonnet 5.0 is gone from the model picker. Opus Plan Mode, chat compaction and background routines also use Sonnet 5.5 now.

Images in Xeref Code: image models now draw in Xeref Code sessions. Pick Nano Banana 2 or GPT Image 2 and ask for an infographic: the image appears in the session and is still there after you refresh. Before, the model answered that it could not see images.

v6.2 — Dots, Receipts, and a Skill Page

Xeref Dots (Ultra): create a Dot, an agent that owns an ongoing responsibility. Pick a schedule and the tools it may use (web, your tasks and documents, GitHub, Gmail) and it leaves a report each run, with a full activity timeline. A Dot can only read; anything it wants to remember waits for your approval. The sidebar now has Dots at the top, and xeref.ai/dots opens the page directly.

Answer receipts: every chat reply now shows which model answered, how many tokens it read and wrote, and how fast it started.

Agent Skill page: xeref.ai/skill has a skill you can copy into Claude Code, Cursor or any MCP client, so your agent can read and update your projects, tasks and notes. Agents can also fetch it as plain markdown from xeref.ai/skill.md.

Models: Claude Sonnet 5.5 joins the picker for Pro and Ultra, with stronger reasoning at Sonnet speed. GPT-6.1 Sol, OpenAI's latest flagship model, is available on Ultra. The pricing page lists them and has a new Developer mode section for connecting your agents.

Plan limits enforced on the server: saving, MCP and IDE access, deploy channels, scheduled workflows, documents and OCR, and Xeref Code are now checked against your plan on the server, not just in the interface. If you downgrade, nothing is deleted: existing data stays readable, new items are blocked, and your oldest channels and scheduled jobs inside the limit keep running.

Housekeeping: short links (xeref.ai/dots, /agents, /claw, /hermes, /jev and /harness) open the matching page, agent names are consistent everywhere (Xeref Claw, Xeref Hermes, Xeref Jev, Xeref Harness), and adding the starter routines no longer fails with a duplicate-key error.

v6.1 — MCP Tools in Dispatch

Tools that work: in Settings → Dispatch, switch on Context7 and chat can look up current library docs. GitHub and Supabase let it read your repos, issues, pull requests and projects (read-only, and each needs its connection under Customize → Connectors). Sequential Thinking lets it reason through a hard problem step by step. They are available on Pro and Ultra, and every server starts switched off.

A shorter, honest list: servers that were listed but never connected to chat are gone, and a server that still needs a key or a connection says "Setup needed" instead of showing a switch that does nothing.

v6.0 — Open in Code

Task to session: open any task in Xeref Code and you get a session linked to it, with a short context brief and a proposed implementation plan.

Review before work starts: approve the plan, reject it, or request a revision with feedback. Every decision is kept in a session timeline. When the work is done, record a completion summary (what changed, which checks passed, follow-ups) and choose whether to move the linked task to its new status. A new Code sessions page lists your sessions by status, project and linked task.

Chat questions: the click-to-answer questions from v5.9 gained an "Other…" choice so you can type your own answer right in the question, and they keep their buttons after you refresh or reopen a chat.

The through-line of v6: make the handoffs visible. A plan you approve before code is written, a receipt that shows what answered you, a Dot that can only read and asks before remembering, and plan limits that hold no matter where the request comes from.