ScreenCue 0.2.0 - Native AI client integrations

Native client integration packages for Codex, Claude Code, Claude Desktop, Gemini CLI, and Cursor, with English documentation, runtime labels, examples, and previews.

• Codex: portable plugin, compatibility manifest, local marketplace, MCP configuration, and skill.
• Claude Code: native plugin, MCP configuration, and skill.
• Claude Desktop: native .mcpb extension and a separate uploadable skill ZIP, for Windows/macOS.
• Gemini CLI: extension, MCP configuration, skill, context, and show/hide/status commands.
• Cursor: local plugin, MCP configuration, skill, and scoped rule.
• Optional Python installer with safe configuration merges, backups, dry run, and a connection doctor. Packaged plugins run the native binary directly, without Python or Node.js.
• Agent guidance prioritizes passive observation and immediate pointing, without opening another window or relaying the request to another chat. Hiding a cue calls screen_clear directly.

Download: choose screencue-integrations-PLATFORM.zip for ready-made client packages, or the smaller screencue-PLATFORM archive for the standalone binary. Extract into a permanent folder and follow the included README or client guides. Use one connection per client to avoid duplicate overlays.

Validation: 44 Rust unit tests, 15 installer/package tests, MCP protocol checks, and the Windows native overlay suite passed. All four CI builds passed. The downloaded release archives were checked for matching platform binaries, executable permissions, licenses, resolved plugin paths, and English text. Actual Codex plugin loading/tool calls, Claude Code plugin MCP connection, and Gemini CLI MCP connection were tested. Claude Desktop MCPB and Cursor plugin format validation passed. See the detailed verification record for exact versions and evidence.

Known Gemini CLI issue: with Gemini CLI 0.61.0 and Node.js 24.12.0 on the tested Windows machine, extensions link registered the extension but then failed during CLI shutdown. A separate unattended extensions install attempt timed out. Skill discovery and MCP connection still passed. The repository's standalone scripts/integrate.py configure --client gemini route avoids extension installation and its configured MCP connection passed the doctor check. This is a client lifecycle issue, not a claim that every installation path is clean.

These packages do not supply screen vision or guarantee authenticated visual interaction in every AI client. Direct Dot voice access remains unverified. macOS and Linux graphical desktop behavior still requires native manual testing; Linux requires X11 with a compositor, and native Wayland is unsupported. Binaries and MCPB packages are unsigned.

Hide a hint with Ctrl+Alt+X (Ctrl+Option+X on macOS). Existing click-through drawing, adaptive red, animated sketch arrow, independent target outlines, and local window tracking remain available.

Checksums are in SHA256SUMS.txt. Code is MIT licensed; embedded IBM Plex Mono is SIL OFL 1.1. Third-party license notices are included.

ScreenCue 0.0.4 - Desktop app and independent model connections

ScreenCue now opens as a native desktop application with dark settings, a red sketch icon, and independent conversation notes.

Downloads

• Windows: ScreenCue-0.0.4-windows-x64-setup.exe — per-user installer, Start menu entry, optional desktop shortcut, uninstall.
• macOS: the Apple Silicon or Intel DMG — drag ScreenCue into Applications; macOS 14.2+ (system audio requires 14.6+).
• Linux: the X11 app archive — extract and run sh install.sh for a user-menu installation.
• Smaller renderer-only archives and Codex/Claude/Gemini/Cursor integration bundles remain available.

What's included

• Independent screen_prompt MCP tool: display speaking notes immediately without screen coordinates or a screenshot.
• Settings for appearance, adaptive text-panel width, animation, capture privacy, and explicit conversation Start/Stop.
• Four independent connection profiles: Audio & speakers, Fast model, Big model, and System 1. Each supports its relevant API or command-bridge contract, model selection, and its own credential. Keys can use Windows Credential Manager, macOS Keychain, Linux Secret Service, or an environment reference; they are not saved in settings JSON.
• Windows native capture exclusion for the settings and overlay windows, with exact policy readback. Capture compatibility is application-dependent; macOS/Linux report unsupported.
• Optional native microphone/system-audio helper, OpenAI-compatible endpoints, generic command adapters, structured speaker IDs, and explicit role mapping.
• Optional independent fast LLM, slow planning, and typed-decision lanes. Jev and other configured decision adapters can act without waiting for either LLM. The planner prepares bounded response/question/topic-bridge candidates; fast text can stream while it works.
• Stale-answer rejection, bounded queues, safe action allowlists, stop cancellation, and coordinated hiding across active connections. Ctrl+Alt+X hides cues and stops listening.

No audio starts automatically. Models, accounts, API keys and model weights are not bundled. The native app/basic audio helper need no interpreter; the optional advanced orchestrator requires an existing Python 3.11+ executable configured by absolute path.

Verification and limits

Local verification includes 86 native app tests, 12 native helper tests, 85 passing Python tests, a synthetic Windows credential-store round-trip, live MCP hints, native privacy/style checks including minimize/restore, and UI Start/Stop/restart using a synthetic helper. Provider tests use fake HTTP/command/SSE responses; no live meeting audio or paid model service was used.

Release CI passed on all four platforms, including Windows native UI and installer install/uninstall, macOS bundle signing and DMG verification, and Linux user installation/removal. Downloaded assets were collected from this run for commit 61d98ab3f5ee0c32da13d5b4fcf18919e0c1ad96; SHA256SUMS.txt covers all twelve packages.

Stable speaker identity requires a suitable diarization adapter. Live diarization accuracy and speech-to-useful-cue latency are not benchmarked. Universal screen-share invisibility, native Wayland, and arbitrary exclusive-fullscreen game support are not claimed.

The Windows installer is unsigned. macOS uses ad-hoc integrity signing, not Developer ID notarization.

Setup · Connections and keys · Conversation architecture and providers · Verification

ScreenCue 0.0.3 - Adaptive text panels

This prototype adds independent, content-sized text panels in the selected dark brush style.

• Separate panel position, bold title, readable body, and panel-only mode.
• Up to 16,000 body characters, complete-line pagination, and page geometry in MCP/JSON replies.
• Cached panel rendering, improved arrow avoidance, and fewer redundant render passes.
• Optional Python controller in the source repository: TypeSafe Jev, OpenAI Decisions, Ollama Choice, and a command adapter for other agents/models. It selects supplied observed candidates, applies deadlines, and discards stale answers. It does not add screen capture or a model runtime to the native executable.
• Updated native integration packages for Codex, Claude Code/Desktop, Gemini CLI, and Cursor.
• Explicit presentation diagnostics and X11 composition hints.

Verification: 56 Rust tests, 43 Python tests, live MCP checks, Windows native regression checks, and a controller-to-native JSON smoke test. Provider tests use local fixtures; live model accuracy and latency have not been verified. See the verification record.

All four CI builds passed: Windows x64, macOS Apple Silicon, macOS Intel, and Linux x64/X11. The downloaded Windows MSVC artifact also passed the native and live MCP suites. Its executable is 2,739,712 bytes (2.61 MiB), SHA-256 d06012dfefd756f1c898693fcd1cec19edd8728c3972bb26834a0828b18a2f25. The release archives and integration packages come from that CI run; SHA256SUMS.txt covers all eight archives.

True exclusive fullscreen is not universally supported. No specific game was tested. Native Wayland, Game Bar, and in-game rendering adapters remain future work. Native hide does not yet cancel pending optional-controller inference; controller clear does.

The version number follows the new prototype policy: 0.0.x, with letter suffixes for small follow-up revisions. Earlier 0.1.0 and 0.2.0 releases remain unchanged; this is not a feature rollback.

Download the native archive for your platform. Integration archives contain client-specific plugins and configuration helpers. The optional controller is available in the source archive/repository and requires Python 3.11+. The native overlay needs no Python, Rust, Node.js, or embedded browser to run.

ScreenCue 0.1.0 - Agent hints over your desktop

First preview of ScreenCue: an agent can point to a control directly over the current desktop.

• Large animated SKETCH arrow with four hand-drawn frames and adaptive contrast that prefers red.
• Independent ellipse, rectangle, rounded rectangle, or polygon target outline, with a gap before the arrow tip.
• Click-through hints that retain no keyboard focus and remain until cleared or hidden with Ctrl+Alt+X (Ctrl+Option+X on macOS).
• Local tracking of the original window, with a labeled edge cue when the target is offscreen or the window is minimized. Scrolling, layout changes, and resizing require refreshed coordinates.
• Local stdio MCP and newline-delimited JSON for other agents. Screen observation and voice belong to the host client.
• Native binary with no Rust, Node.js, .NET, or browser runtime required.

This is an older preview. Use the newest release for English runtime labels and the complete native client integration packages.

Download your platform archive, extract it, and run screencue demo (.\screencue.exe demo in Windows PowerShell). See the integration guide.

Windows x64; macOS Apple Silicon or Intel with macOS 12.3+; Linux x64 with X11 and a compositor. Local macOS color sampling requires macOS 15.2+ and existing screen-recording permission. Native Wayland is not supported. Protected desktops and some fullscreen applications may limit overlays.

All four builds passed CI. Live desktop and MCP checks were performed on Windows; macOS/Linux desktop interaction and each AI client require separate validation. Binaries are unsigned. SHA-256 checksums are in SHA256SUMS.txt. Code is MIT licensed; IBM Plex Mono is SIL OFL 1.1.