Changelog

// Release notes and updates for SuperBased

v2.0.13

Latest April 2026

ffmpeg-static cross-arch fix for Intel macOS

  • Intel macOS dictation fixed — ffmpeg-static ships one binary per install selected by host arch. Cross-installed both Darwin x64 + arm64 binaries during CI build, and the runtime resolveFfmpegPath now picks the arched sibling at ${path}-${process.arch} with fallback. Spawn errors with exit=null on Intel Macs are gone.
  • asarUnpack expanded for **/ffmpeg-static/* so all three binaries (ffmpeg, ffmpeg-x64, ffmpeg-arm64) land outside the read-only archive where they're actually executable.
  • Each .dmg ships ~45MB heavier (bundled wrong-arch ffmpeg). Acceptable cost for not breaking ~20% of the Mac user base on first launch.

v2.0.0

Major release April 2026

GUI Automation Suite (S0–S9 + Track C) + Humanization v2

  • 72 MCP tools (was 28) — full agent-hands surface across capture, dictation, and the entire GUI automation suite.
  • GUI primitives (13) — click, type, hotkey, scroll, drag, hover, pixel_color, mouse_position, wait, wait_for, locate, ui_dump, accessibility_tree. Reliability pyramid: AutomationId / AXIdentifier → role+name → OCR label.
  • Sequence orchestration — superbased_sequence bundles N steps under one approval, one activation. Per-step returnObservation; explicit screenshot steps return PNG inline.
  • Semantic + macros (5) — ax_invoke (UIA Invoke / Toggle / SelectionItem.Select / Value.SetValue / ExpandCollapse), form_fill, dialog_handle, context_menu_select, drag_file (Windows scaffold).
  • Window + browser control (8) — window_state, window_bounds, resize_window (with device presets), focus_window, launch_app (with sha256 hash pinning), open_url, tab_management, find_in_page.
  • Specialized (3) — tray_click (Shell_TrayWnd cross-process reads), virtual_desktop, doctor_gui_automation.
  • Cross-platform parity — full Windows + macOS via S9 + Track C. macOS routed through nut-js CGEventPost + osascript System Events AX + @superbased/macos-ax binding + /usr/bin/open.
  • Humanization v2 — per-call humanize: 'off' | 'light' | 'human' | 'paranoid'. Bezier-curved cursor approaches with sin-shaped velocity envelope, gaussian click-target jitter, gamma-distributed inter-key timing, click + key hold variation (50–110ms vs ~5ms atomic), pre-click settle dwell + tremor, optional typo+correct sequences. Per-process cross-session salt mixed into seeds. Defeats trajectory and timing classifiers used by reCAPTCHA, Cloudflare Turnstile, hCaptcha, GeeTest, KeyCAPTCHA.
  • Safety rails — master toggle guiAutomation.enabled (off by default); per-action group toggles (safe / write / destructive); confirm flag required; protected-apps blocklist; protected-process geofence (SELF_TARGET_REFUSED); kill switch (Ctrl+Shift+Esc).
  • Audit log + replay — every action recorded as NDJSON to ~/.superbased/audit.log. superbased_replay reproduces trajectories byte-for-byte within a process.
  • Stdio bridge co-existence — superbased mcp detects a running desktop instance via PID file and acts as a stdio↔HTTP bridge to its :47592 endpoint. Stdio MCP clients work whether the desktop, a headless serve, or nothing is running.
  • Plugin ecosystem — 7 platforms shipped: Claude Code, OpenAI Codex, Cursor, GitHub Copilot CLI, Cline (MCP marketplace), OpenCode (npm), OpenClaw (ClawHub).

v1.6.7

Stable April 2026

Dedicated multi-region capture hotkey + predictable re-press behavior

  • Multi-region hotkey — new Ctrl+Shift+D (Cmd+Shift+D on macOS) opens the selection canvas in forced multi-region mode. Every drag accumulates a region without needing to hold Ctrl. Fully customizable in Settings > General.
  • Done button & Space to confirm — floating "Done (N)" button bottom-right once you've selected at least one region, or press Enter or Space to commit all regions at once. Right-click pops the last region.
  • Appends to open overlay — if the overlay is already open, multi-region captures are added to the current session as separate thumbnails, matching the single-region "add to session" behavior.
  • Predictable re-press — pressing any capture hotkey (Ctrl+Shift+S / D / A) while the selection canvas is already open now cancels it instead of reopening a second window. Double-tap to dismiss.
  • Stale state fixes — scroll-capture mode and "adding to overlay" flags now reset on every selection close path (complete, Esc, crash), eliminating mode leaks between captures.
  • Flicker fix — right-clicking to pop a region no longer flashes the canvas. Canvas sizing is now decoupled from region state.

v1.5.0

Stable April 2026

Headless mode, MCP improvements, dictation performance

  • Window targeting — capture any window by name, even minimized (ShowWindow + SetForegroundWindow via Win32 P/Invoke)
  • 28 MCP tools + 9 resources — expanded tool descriptions, skill docs rewritten for better agent discovery
  • Dictation cleanup pipeline — full 4-stage processing: deduplication, technical term normalization, filler removal, short phrase handling
  • 100+ brand name normalizations — "Cloud Code" → Claude Code, "Super Bass" → SuperBased, "deep seek" → DeepSeek, and more
  • 955 Azure Speech phrases — 170 new AI/ML terms (models, frameworks, tools, vector DBs)
  • Dictation performance — clipboard pre-save moved to recording start (~200-300ms faster stop path)
  • MCP tool descriptions — window targeting, dictation mic recording, clipboard readImage documented
  • superbased mcp command — unified API server + MCP stdio bridge in one process
  • Published to npm as superbased@1.5.0

v1.4.0

Stable April 2026

Headless CLI + MCP server, push-to-talk dictation

  • Headless mode — npm global package, no Electron dependency, full API server + MCP bridge
  • superbased mcp — single command for all AI editor integrations
  • Push-to-talk dictation — Cmd+Opt (macOS) / Ctrl+Alt (Windows) hotkey, warm mic, standby Azure recognizer
  • Window enumeration — Win32 EnumWindows via C# for multi-window apps
  • Capture hotkey — Ctrl+Shift for instant fullscreen screenshots
  • Free tier — 50 AI credits/day for first 30 days
  • MCP configs for Claude Code, Cursor, Windsurf, Cline, OpenCode, Zed, Codex

v1.0.0

Stable March 2026

Initial public release

  • Region, window, fullscreen, scroll, and timed captures
  • Image editor with annotations, blur, redact, numbered steps
  • AI analysis with OpenAI, Anthropic, and Google AI (BYOK)
  • Workflow recording with visual regression testing
  • Gallery with search, tags, notes, and statistics
  • OCR text extraction
  • Pin screenshots to desktop
  • Color picker tool
  • Auto-updates via GitHub Releases