Back to Blog
May 16, 20262 min read

Skills, Voice, and Going Open Source: Xeref v2.6–v2.9

XerefAIVoiceOpen SourceRelease Notes

Four back-to-back releases that add a Skills System, customizable sidebar, F9 voice dictation, and move the xeref repo to public with full branch protection and CI.

May 13–16 was a rapid-fire stretch of releases that each added something distinct: extensibility, personalization, voice, and transparency.

v2.6 — Skills System

A full Skills System landed with database migration, server actions for CRUD, and new UI components (SkillsSection, SkillFileTree, SkillContentPane) under /customize/skills. This is the foundation for packaging reusable instruction sets that can be shared across agents and with the Skool community.

MCP Server memory tools were also renamed and enhanced — list_memories, save_memory, recall_memories, delete_memory — aligning with the CLAWS Archive methodology. An OAuth redirect bug affecting Google, Notion, Slack, and Vercel flows was resolved at the same time.

v2.7 — Customizable Sidebar

The sidebar is now user-configurable: toggle visibility, reorder nav items, and persist preferences to your profile. Hidden items collect under a "More" group. This release also introduced profiles.preferences — a JSONB column that backs sidebar config, capabilities settings, and Xeref Code settings across all future settings tabs.

A meaningful infrastructure fix shipped too: memory file uploads now bypass Vercel's 4.5 MB serverless payload limit by going directly to Supabase Storage via signed upload URLs. Files up to 50 MB upload without error.

v2.8 — Open Source and CI

The xeref repo is now public. Main branch protection is in place — PRs require 1 approval, passing CI (Lint, Type Check, Tests), and the branch is guarded against force pushes and deletion via a GitHub Ruleset. CI reliability was improved by excluding the agent/ folder from the root tsconfig and bumping actions/checkout and actions/setup-node to v5.

v2.9 — Voice to Text

Press F9 anywhere in the dashboard to start or stop voice dictation. Transcribed text is appended to the chat composer. A microphone button also lives next to the send button for click-to-dictate. Recordings are transcribed by Groq whisper-large-v3-turbo and lightly polished for grammar without paraphrasing.

Four releases, four directions — each one removing a friction point that had been on the list since early beta.