xeref.ai logo
xeref.ai
Changelog

What's new

Release history for xeref.ai — from newest to oldest.

v4.8

LatestJune 13, 2026
New
  • Per-agent skills — XerefClaw, Xeref Agents, and XerefHermes now each have their own skill set instead of sharing one toolbox. XerefClaw keeps the full 9 skills (tasks, projects, memory, documents, web), Xeref Agents focuses on research (memory, documents, web), and XerefHermes specialises in memory and knowledge. Ask any agent "what skills do you have?" or request a comparison between agents and the answer now comes from real configuration, not guesswork.
  • Built-in Agent Skills matrix — the Skills page (/customize/skills) now shows exactly which skills each system agent has, grouped by category, and the agent picker in chat displays a skill count for every agent.
  • How Xeref Works — the About page has a new interactive showcase that walks through the product story (Build, Orchestrate, Deploy) with a scroll-linked visual that follows you as you read.
  • GitHub organization visibility — the GitHub connection now requests organization read access and shows which orgs your token can reach next to "Connected as", with a clear warning when no organization has granted the app access.
Improved
  • Supply-chain and secret hygiene — every push now runs an automated secret scan across the full git history, dependency vulnerability alerts and automated security fixes are enabled, and the test suite now enforces a code-coverage floor in CI.
  • Pricing page refresh — the Ultra plan now leads with Claude Fable 5 and the Ultra model lineup is complete: DeepSeek V4 Pro is now included, so Ultra is a true superset of Pro.
  • Referral program copy — eligibility wording is now consistent everywhere: rewards count when your friend signs up through your link and creates their first agent project.

v4.7

June 12, 2026
New
  • Text transforms in the chat composer — press Ctrl+. (⌘+. on Mac) to polish what you wrote into a clearer, cleaner message that keeps your voice, or Ctrl+Shift+. to turn rough notes into a structured, AI-ready prompt with role, task, and context sections. Transforms apply to just the selected text, or the whole message when nothing is selected, with a live indicator while the rewrite runs. Both shortcuts are documented in the keyboard shortcuts dialog.
  • MCP Servers settings (Xeref+) — Agent Manager settings has a new MCP Servers tab: every server in the catalog with its description and tool count, a details pane showing the exact tools a server exposes, and one-click enable/disable that stays in sync with the dispatch panel. Two new servers join the catalog: Notebooks (internal workspace notebooks and cell operations) and Visualization (Apache ECharts v5 chart rendering via render_chart).

v4.6

June 10, 2026
New
  • Claude Fable 5 — Anthropic's new Mythos-class model is now available on Ultra. Select "Fable 5" in the chat model picker for the deepest planning and long-horizon reasoning available on the platform (1M token context). Opus Plan Mode now routes planning-shaped queries to Fable 5 automatically, and the Plans generator uses Fable 5 for Ultra accounts.
Improved
  • Plan generation resilience — if the primary model declines or returns unusable output, plan generation now automatically retries on Claude Sonnet 4.6 instead of failing, and output is capped to keep generations fast and bounded.

v4.5

June 9, 2026
New
  • Builder feature catalog expanded from 22 to 48 — the /builder (XerefClaw) catalog now offers 48 agent capabilities across all six CLAWS categories. New entries: Connect (SMS via Twilio, Microsoft Teams, Web Chat Widget), Listen (OCR, Video Understanding, Speaker Diarization, Language Translation), Archive (Knowledge Graph, SQL Memory, Entity Memory, Hybrid Search), Wire (Notion, Google Sheets, Stripe, HTTP/REST Tool, Linear/Jira, Outbound Webhooks), Sense (Webhook Trigger, RSS Monitor, Inbox Watcher, Website Change Monitor, Reminder Scheduler), and Agent Architecture (Reflection, ReAct Tool Loop, Guardrails, Structured Output). All prompts scaffold what the generated agent can do today.

v4.4

June 8, 2026
New
  • Download page — xeref.ai/download is now a real page explaining how to continue your workspace on any device. Three-step "how it works" walkthrough, a list of everything that travels with your account (projects, agent chats, Pinecone memory), and an honest roadmap card for the native mobile app. xeref.ai/download#mobile deep-links directly to the mobile section.

v4.3

June 7, 2026
New
  • Read any URL — XerefHermes can now fetch and read the text content of any web page you share. Paste a URL into chat and ask it to read, summarise, or extract information from the page. Available on all plans.
  • Image analysis — Attach an image or paste an image URL in chat and XerefHermes will describe, analyse, or answer questions about it. Messages with images are automatically routed to a vision-capable model (Gemini 3.5 Flash on Pro, Claude Sonnet 4.6 on Ultra). Pro and Ultra only.
Improved
  • Document recall security — Uploaded files are now scanned for secrets (API keys, tokens, recovery codes, private keys) before being indexed or returned in chat. Sensitive content is masked as [REDACTED] and files with names like recovery-codes or .env are withheld from recall entirely. Existing documents are cleaned on the next recall.

v4.2

June 3, 2026
New
  • XerefHermes nightly digest — a new dream:summarize handler gives XerefHermes its own "dreaming function". Every morning it reviews the last 24 hours of your tasks and activity, writes a short overnight digest with a suggested next focus, and saves it into long-term memory. Inspired by the Nous Hermes agentic OS.
  • Nightly dream cron — a dedicated /api/cron/hermes-dream route runs daily at 08:00 UTC (11:00 GMT+3), enqueues a dream:summarize envelope for every active user, and dispatches it in the same run. Digests appear as searchable memory entries tagged dream-digest.

v4.1

June 2, 2026
Improved
  • Model picker now shows a short description under each model name — visible in the chat input, agent panel, and Xeref Code. Your last-used model and planning-mode preference are now remembered across sessions.
  • Checkout errors now surface as in-app toasts instead of silent failures — misconfigured keys, network errors, and missing product IDs all show a clear message.
  • MCP server endpoint is fully JSON-RPC 2.0 compliant: every response echoes the request id, notifications return 202, unknown methods return -32601, and tool errors return isError: true instead of HTTP 400. Fixes compatibility with Claude Desktop and strict MCP clients.
  • grok-3 removed from the model list — the model was retired upstream. grok-4.3 and grok-3-mini remain available on Pro and Ultra.

v4.0

May 31, 2026
New
  • Dispatch page — the /dispatch route is now a live settings panel instead of a placeholder. Control agent runtime permissions (keep awake, mobile notifications, browser action policy) and toggle individual MCP tools from a dedicated page.
  • Classroom default curriculum — First-time visitors now see a fully seeded 15-course catalog covering onboarding, core CLAWS methodology, agent architecture, integrations, and monetization. No manual content setup required; the catalog seeds automatically on first load.
  • Connector token health-check — The Connectors page (/customize/connectors) now performs a live API call for each connected service on load. Expired or revoked tokens are flagged in real time so you know exactly which connections need re-authorisation.
  • Artifact lifecycle — Draft artifacts can now be published. Publishing persists the artifact to the database and triggers workflow generation automatically. The artifact panel reflects the published state immediately.
  • Schema consistency: updated_at columns and auto-update triggers added to 9 tables that were missing them (modules, routine_runs, documents, code_messages, feedback, brand_audit_actions, messages, telegram_links, telegram_pairing_codes). Performance indexes added on memories, lessons, and modules.
Fixed
  • Chat file & photo attachments now send correctly. Attachments were silently failing at two levels: the storage bucket was missing (files never uploaded) and a non-cloneable URL object in the message payload was crashing the AI SDK state clone. Both are resolved — images and PDFs reach the model as intended.
  • DOCX uploads are now rejected immediately at the upload step with a clear error instead of being accepted, getting stuck in "processing", and failing silently during text extraction.
  • Cron expressions for routines are now validated at creation time (must be a valid 5-field schedule). Previously any string was accepted and would only fail at runtime.
  • Hermes inter-agent messages that fail handler execution now retry with exponential backoff (1 → 2 → 4 → 8 → 16 min) and are permanently marked dead after 5 attempts instead of staying queued forever.
  • Classroom lesson embedding — New and edited lessons are now indexed into Pinecone immediately on save, fixing a silent failure that prevented lessons from appearing in Classroom semantic search.
  • Changelog page: Skool community link corrected from /classroom to /about.
Improved
  • MCP server endpoint is now rate-limited to 60 requests per minute per user.
  • Workflow webhook triggers now accept the secret via the X-Webhook-Secret header (header is preferred; query-param still works for backward compatibility).
  • Settings and session PATCH endpoints now validate the exact fields they accept — unexpected fields are rejected instead of being silently written to the database.
  • Admin authorization for the Classroom is now driven by the ADMIN_EMAILS environment variable instead of hardcoded values, and consolidated into a single SQL function.
  • Pinecone chunk deletion failures are now logged as warnings instead of swallowed silently.
  • Pricing page: AI model counts are now accurate — Pro lists 12 models, Ultra lists all 17.
  • Pricing page: benefit language rewritten across all three plan cards for clarity and impact.
  • Pricing page: CircleCheck icons replace plain checkmarks; excluded feature rows removed from the Basic card.
  • Ultra plan now surfaces live agent monitoring & inter-agent messaging (Dispatch + XerefHermes) as a visible benefit.

v3.9

May 29, 2026
New
  • Claude Opus 4.8 — our top Ultra model is now Anthropic's newest flagship. Better judgment and code review (~4x less likely to miss flaws), stronger agentic decision-making, and improved alignment at the same price as Opus 4.7. Opus Plan Mode now plans with 4.8 too.

v3.8

May 28, 2026
Improved
  • Navigation performance — Dashboard view switches now trigger a single server round-trip instead of one per data source. Auth verification happens once per navigation, cutting Supabase Auth API calls by ~75%.
  • View data cache — Switching back to a previously visited view (Tasks, Stats) is now instant. Cached data renders immediately; a background refresh keeps it current.
  • Public page layout — About, Changelog, Pricing, FAQ, Privacy, Terms, and Docs share a sticky header and footer via a shared route group. Per-page redundant nav removed.
Fixed
  • Dashboard links to public pages (Changelog, Pricing) now open in a new tab, preserving dashboard state.

v3.7

May 28, 2026
New
  • XerefHermes — Inter-agent message bus for routing typed envelopes between system agents. Any server-side code can publish to the Hermes queue; registered handlers process messages asynchronously.
  • memory:save handler — First Hermes handler: publishes payload.text to the Pinecone xeref_user_memory namespace for the authenticated user. Dispatched by Vercel cron every minute or manually via POST /api/hermes/dispatch.
  • Hermes Inspector — New sidebar view in Code tab → AI Agents → XerefHermes. Shows queued / processing / done / error message counts and a live log of the last 50 envelopes with per-row status icons.
  • XerefHermes System Agent — Added to the System Agents selector in the chat input (Workflow icon). Activates a persona that understands the Hermes envelope schema and available handlers.
Fixed
  • Chat infinite loop — Resolved "Maximum update depth exceeded" error that fired on every message submit. Root cause: ChatsView sync effect had activeChat in its dependency array, causing repeated setActiveChat calls when a new chat was created. Fixed by reading activeChat through a ref so the effect only re-runs when the parent chat list changes.

v3.6

May 23, 2026
New
  • Model Selector — 6 new models added to the chat selector with Pro and Ultra plan badges.
  • Pro models: GPT 5.5 (OpenAI), Kimi K2.6 (Moonshot AI), GLM 5.1 (Z.AI), Grok 4.20 (xAI).
  • Ultra models: GPT 5.5 Pro (deep reasoning, 5–10 min think time), Gemini 3.1 Preview (Google reasoning preview).
  • MCP Setup — Login page now shows a two-step mcp-remote install guide (npm install + JSON config) compatible with Claude Desktop, Cursor, Windsurf, and all stdio MCP clients.
  • MCP Quick Setup in Settings — The MCP Server Token card now includes an inline install command and config JSON block with your token pre-filled, plus a one-click Copy config button.
  • IDE/CLI Brand Icons — Cursor, Windsurf, Claude Code, Antigravity, and Codex CLI tiles on the login page now show accurate brand marks instead of placeholder shapes.
  • Referral Link — Referral URL is now always xeref.ai/login?ref=xeref-pro regardless of the host the dashboard is opened on.
  • MCP Endpoint — GET /api/mcp now returns a friendly JSON info response instead of a browser 405 error.

v3.5

May 22, 2026
New
  • Agent Sessions Panel — A new "Sessions" section in Agent Manager lists all Anthropic Managed Agent sessions with live status dots and pass/fail badges derived from outcome evaluations.
  • Session Detail View — Clicking a session shows a full detail panel with timeline (created / started / completed), an expandable Success Rubric, and the raw outcome payload from the Anthropic webhook.
  • Outcomes Rubric — Sessions with a rubric_passed result display a green PASS or red FAIL badge in both the sidebar and the session detail header.
  • Anthropic Webhook Hardening — The /api/webhooks/anthropic handler now verifies HMAC-SHA256 signatures (Anthropic-Signature header, t=…,v1=… format) with a 5-minute replay-attack window. Events are upserted to a new agent_sessions table via service role.
  • agent_sessions Table — New Supabase table stores session_id, status, rubric, rubric_passed, raw_outcome, started_at, completed_at. RLS restricts reads to the owning user; webhook writes via service role.
Improved
  • Webhook Reliability — Stale payloads (older than 5 minutes) are rejected before signature check. DB errors are logged but never swallow the 200 response to Anthropic.
  • Agent Manager Sidebar — Sessions section is rendered below Brand Tools; empty state is shown until the first webhook arrives, keeping the UI clean for new accounts.

v3.4

May 22, 2026
New
  • Live Conversations — Agent Manager now loads your real chat history from Supabase. Clicking a conversation fetches and displays the full message thread on demand.
  • Live Artifacts — The Knowledge panel is backed by real artifact data. All documents, code, and workflow artifacts from your account appear in the split-pane browser.
  • Live Workspace Stats — Workspace detail cards now derive Total Tasks / Completed / Pending / Active Pipelines / Success Rate from real Supabase tables (tasks + workflows) instead of hardcoded numbers. Workspaces are matched by project name.
  • Feedback Submission — The "Provide Feedback" modal now persists submissions to the database. Bug reports, feature requests, and general feedback are stored with your account, steps to reproduce, and screenshot indicator.
  • Brand Audit Apply Tracking — Clicking "Apply Changes" on the Brand Audit panel now records the applied recommendation to a new brand_audit_actions table for observability and future automation.
  • Anthropic Webhook Receiver — A new endpoint at /api/webhooks/anthropic is ready to receive session.status_* events from Anthropic Managed Agents.
Improved
  • Agent Manager Performance — AgentManagerView is now lazy-loaded via next/dynamic, so it only loads when Ctrl+E is pressed. This reduces initial dashboard JS by the full 1 900-line component bundle.
  • Empty States — Conversations and Artifacts panels show a subtle loading indicator while fetching, and a clear empty-state message when no data exists.

v3.3

May 21, 2026
New
  • Screenshot Uploads — Attach screenshots directly in the Provide Feedback modal via a functional file picker with inline preview and memory-safe cleanup on close.
  • Chat Input Focus Ring — Cyan glow on the chat input area provides clear visual focus state, improving keyboard navigation and reducing input ambiguity.
  • Launch Configurations — `.claude/launch.json` ships with 8 dev server configurations auto-detected from the workspace, so every project starts with a single command.
  • IDE Integrations Changelog — v3.2 changelog updated with an IDE Integrations section documenting Antigravity, VS Code, Claude Code, and Xeref Code support.
Ecosystem
  • Three-Brain Stack — Bugra Karsli ships xeref.ai on a unified agentic dev stack: Antigravity IDE for day-to-day MCP-powered work, Claude Code CLI for deep cross-file refactoring, and Xeref.ai as the deployment and agent management surface.
  • Antigravity IDE — Primary development environment with native MCP server support and the `.agent/skills/` ecosystem, enabling tool-augmented workflows without leaving the editor.
  • Claude Code CLI — Handles codebase-wide migrations, complex refactors, and multi-file changes that require full repo context — paired with Xeref.ai for agent configuration and deployment.
  • Xeref.ai as the Third Brain — The platform acts as a live agent builder, productivity dashboard, and dispatch layer — closing the loop between code, configuration, and autonomous execution.
Improved
  • Feedback Modal UX — File attachment state is fully cleared on modal close, preventing stale previews or memory leaks across repeated feedback submissions.
  • Changelog Completeness — IDE integration milestones are now reflected in the v3.2 entry, keeping release history accurate as tooling support expanded post-ship.
  • Workspace Dev Ergonomics — Auto-detected launch configs reduce friction when switching between the 8+ active projects in the workspace root.

v3.2

May 21, 2026
New
  • Agent Manager Surface (Ctrl+E) — Ctrl+E now opens a full workspace-aware Agent Manager dashboard instead of just a settings overlay. Features a sidebar with live conversation history, workspace navigator, and Brand Tools section.
  • Brand Audit Panel — A new Brand Audit view scores xeref.ai against the CLAWS methodology (Connect, Listen, Archive, Wire, Sense, Identity) with a live 87% alignment gauge, gap detection, and one-click copy remediation.
  • Provide Feedback Modal — The "Provide Feedback" button now opens a full-featured modal with feedback type selector (Bug Report, Feature Request, Auth & Billing, General), description field, steps to reproduce (bug mode), screenshot attachment, and Antigravity log attachment toggle.
  • Knowledge Base View — A new Knowledge tab under "More" provides a split-pane artifact browser. Each artifact supports Copy Content, Copy Path, and Export Artifact (downloads the file to disk) via a "..." context menu.
  • Chrome Remote Debugging Safety Popup — Navigating to Browser settings now shows a security advisory popup explaining the risks of remote debugging, with a dedicated debug Chrome profile recommendation and one-click launch command copy.
  • Permissions & Policies Settings Tab — New "Permissions" tab in Settings exposes four security toggles: Enforce Local Rules, Terminal Execution Sandbox, Restricted Network Boundary, and Artifact Modification Policy.
  • Conversation Thread View — Clicking any conversation in the sidebar opens a full message thread with role-based bubbles, inline reply composer, and voice input support.
  • Workspace Detail Cards — Each workspace entry now expands to show agent count, task completion stats, active pipeline info, and an Open Editor button.
  • Screenshot Upload in Feedback — The "Attach a screenshot" button in the Provide Feedback modal is now fully functional: opens a native file picker, renders an inline preview thumbnail, and supports clearing before submit.
  • Cyan Focus Ring on Chat Input — The chat composer now shows a cyan border glow on focus for clearer visual affordance.
IDE Integrations
  • Antigravity IDE — Xeref is now the native MCP-connected workspace for Antigravity. All agent skills, `.agent/skills/` symlinks, and MCP tool definitions are auto-discovered by Antigravity on project open. CLAWS prompt generation outputs directly into the active Antigravity session.
  • VS Code — Xeref MCP server is now compatible with the VS Code MCP extension (github.copilot.chat.mcp). Add your Xeref MCP endpoint and bearer token to VS Code settings to expose tasks, projects, and memory tools directly in GitHub Copilot Chat.
  • Claude Code — Xeref ships a first-class Claude Code integration. The self-hosted MCP server at `/api/mcp` is pre-wired in `.claude/settings.json`; Claude Code can read and write tasks, projects, and notes without leaving the terminal.
  • Xeref Code — The `/code` workspace has been promoted to the primary coding surface. New session picker, GitHub repo selector, model/effort toggles (Default / Accept Edits / Plan), and full streaming chat are all live. Sessions persist to Supabase and resume on reload.
Improved
  • Agent Manager sidebar — Added Conversations section with unread indicators and status badges (active/idle/completed). Brand Tools section added below Workspaces.
  • Agent preset selector — New dropdown in the conversation composer lets you pick from Code Architect, Test Engineer, Security Auditor, CI/CD Deployer, and RAG Memory Manager.
  • Sign Out button removed from sidebar dropdown — sign-out now lives exclusively in the Settings modal for a cleaner quick-access menu.
  • Escape key propagation — Pressing Escape now closes modals in order: Settings → Shortcuts → Browser Debug popup → Audit Logs → then returns to Editor.
  • Global Ctrl+B shortcut — Toggles the Agent Manager sidebar collapsed/expanded.
Fixed
  • Missing Sparkles icon import causing a runtime crash in the Brand Audit sidebar button.
  • Feedback button was a non-functional stub — now opens the full feedback modal.

v3.1

May 19, 2026
New
  • Scheduled Agents — Set a cron schedule for any agent to run automatically in the background.
  • Dynamic Subagents — Agents can now spawn specialized sub-agents mid-task (e.g., /browser, /search) for parallelized workloads.
  • Agent Artifacts — View and download structured deliverables and outputs directly from agent conversation histories.
  • Skills Library — Package and reuse custom instruction packs across agents, with full shareability for the Skool community.
  • Browser & Dispatch Settings — Configure custom browser automation parameters, runtime permissions, and MCP tool access permissions directly.
Improved
  • Permission Modes — Strict Mode toggle added to agent settings to require human approval before sensitive actions are executed.
  • Usage Dashboard — Gain transparency with Settings view displaying per-model rate limits, credit tracking, and refresh timers.
  • Run History & Diff View — Inspect exactly what files, data, or filesystems each agent touched and modified in a session.
  • Global Settings Shortcut — Pressing Ctrl+E now directly triggers the settings panel and focuses the Browser tab.
Fixed
  • MCP server connection reliability and timeout improvements.
  • Fixed workspace-specific settings not persisting correctly across sessions.

v3.0

May 17, 2026
New
  • Keyboard Shortcuts panel — a ⌨ icon button now lives above the v3.0 badge in the right rail. Click it to open a reference dialog listing every real wired shortcut (Ctrl+1/2/3, Ctrl+E/L, Ctrl+Shift+O, F9, F, Ctrl+B) grouped by category. Mod key auto-renders as ⌘ on macOS.
  • Profile avatar in sidebar — the bottom-left user avatar now mirrors the photo uploaded in Settings → General instead of always showing initials. Falls back to initials when no photo is set.

v2.9

May 16, 2026
New
  • 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.
  • Whisper-grade transcripts — recordings are transcribed by Groq whisper-large-v3-turbo and lightly polished for grammar and punctuation without paraphrasing.
Improved
  • Chat composer — the existing MicButton infra used in the legacy AgentPanel is now available in the main chat view too, with a shared global toggle event so future hotkey surfaces can hook into the same recording.

AI Automations by Bugra Karsli

Learn More
Continue on Phone