OpenParable
Skip to content
← Back to skills

Openclicky Screen Control

Apps & Automation
openclicky-screen-control

Openclicky Screen Control is an Agent Skill for apps & automation published in the jasonkneen/openclicky repository on GitHub: a SKILL.md instruction file an AI agent can load to do this task.

Instantly control OpenClicky's local overlay bridge to point on screen, show captions, or speak through OpenClicky's voice. Use when the user says "show me where you mean", "show me how to", "how do I do this", "point to it", "highlight that", "say this", or asks for visual on-screen guidance.

View on GitHub ↗
QUICK START

How to use this skill

Bring this guide into your coding agent with a prompt tailored to the tool you use.

  1. Open your project in Codex.
  2. Copy and paste the prompt below.
  3. Review files and risks before approving any action.
Prompt to paste
Inspect the Agent Skill at https://github.com/jasonkneen/openclicky/blob/HEAD/AppResources/OpenClicky/OpenClickyBundledSkills/openclicky-screen-control/SKILL.md and all supporting files. Report provenance, permissions, scripts, dependencies, and security risks. Propose an exact installation into .agents/skills/openclicky-screen-control, but wait for my approval before writing files or running scripts.

Copying this prompt does not install or run the skill. Review third-party files before use.

OpenClicky compatibility guardrails

  • Follow ../_shared/OpenClickySkillCompatibilityPolicy.md before acting.
  • Verify required local commands, tools, keys, or bridge endpoints before promising execution.
  • Treat sends, publishes, deploys, deletes, moves, merges, playlist/library changes, cloud writes, and app-control clicks as external writes unless this skill narrows them further.
  • Stop and report the exact missing setup step for unavailable tools, auth, or macOS permissions; do not loop or silently switch to browser automation.

Use OpenClicky's local external-control bridge for immediate visual guidance. Do this instead of describing where to look when the user asks you to show them.

Transport

Local-only HTTP/SSE bridge:

http://127.0.0.1:32123

Health check:

curl -s http://127.0.0.1:32123/health

The bridge is designed to be fast and non-invasive: commands only drive OpenClicky's proxy overlay/voice layer and do not start dictation, submit prompts, create agent sessions, or mutate the main app conversation.

Important cursor model:

  • OpenClicky's cursor is the little triangle that normally follows the user's real system pointer.
  • Default /cursor uses OpenClicky's native smooth pointing choreography: the triangle zips to the target, captions it, then flies back. It does not warp the real macOS pointer and does not draw a duplicate primary cursor icon.
  • Use mode: "secondary" only when you intentionally want an extra temporary colored pointer. Secondary pointers disappear automatically.

Coordinates

Use macOS/AppKit screen coordinates: origin at the bottom-left of the global desktop. Use screenshot context, window geometry, or visible UI positions to estimate points. Accuracy matters less than relevance: only point when the target is clearly connected to the user's current request. Never point at an unrelated control just to provide a visual cue.

Fast commands

Point at a location with the primary cursor

curl -s -X POST http://127.0.0.1:32123/cursor \
  -H 'Content-Type: application/json' \
  -d '{"x": 640, "y": 520, "caption": "Click here", "durationMs": 4500}'

Optional fields:

  • caption: short text shown beside OpenClicky's primary cursor
  • durationMs: how long to keep the caption visible; default is about 4s
  • travelMs: accepted for compatibility, but primary cursor motion uses OpenClicky's native smooth pointing choreography.
  • accentHex: caption color, e.g. #60A5FA
  • mode: default primary; use secondary only for an extra temporary pointer

Show extra temporary cursors

Use this when showing multiple possible locations or comparing alternatives. These are additional colored OpenClicky-style cursors, not the user's primary pointer.

curl -s -X POST http://127.0.0.1:32123/cursors \
  -H 'Content-Type: application/json' \
  -d '{"cursors":[{"x":640,"y":520,"caption":"Option A","accentHex":"#60A5FA"},{"x":900,"y":520,"caption":"Option B","accentHex":"#34D399"}],"durationMs":4500}'

Or a single secondary cursor:

curl -s -X POST http://127.0.0.1:32123/cursor \
  -H 'Content-Type: application/json' \
  -d '{"x": 640, "y": 520, "caption": "Look here", "mode": "secondary"}'

Show a caption near a point

curl -s -X POST http://127.0.0.1:32123/caption \
  -H 'Content-Type: application/json' \
  -d '{"x": 900, "y": 700, "text": "This is the setting you want", "durationMs": 5000}'

If x/y are omitted, OpenClicky shows the caption near the current mouse location.

Capture screenshots to locate something

When you need to find something before showing it, request screenshots first. The response includes local JPEG paths and display frame metadata in the same AppKit coordinate space used by /cursor.

curl -s -X POST http://127.0.0.1:32123/screenshot \
  -H 'Content-Type: application/json' \
  -d '{"focused": false}'

Use focused: true to capture only the focused window when possible.

Workflow: capture screenshot → inspect/recognize target → call /cursor with a short caption.

Speak without entering voice mode

curl -s -X POST http://127.0.0.1:32123/speak \
  -H 'Content-Type: application/json' \
  -d '{"text": "Click the button in the top right."}'

If OpenClicky is already speaking, this returns HTTP 409 unless interrupt: true is passed. Prefer not to interrupt unless the user explicitly wants the new instruction spoken now.

Clear the proxy overlay

curl -s -X POST http://127.0.0.1:32123/clear

MCP-style tool endpoints

List descriptors:

curl -s http://127.0.0.1:32123/mcp/tools

If your runtime wants one generic tool-call shape, use:

curl -s -X POST http://127.0.0.1:32123/mcp/call \
  -H 'Content-Type: application/json' \
  -d '{"tool":"openclicky_point","arguments":{"x":640,"y":520,"caption":"This one"}}'

Supported tool names:

  • openclicky_point (preferred single-target pointing tool)
  • openclicky_point_many (preferred multi-marker pointing tool)
  • show_cursor / openclicky_show_cursor (compatibility aliases)
  • show_cursors / openclicky_show_cursors (compatibility aliases)
  • show_caption
  • screenshot
  • speak
  • clear

For multiple coordinated tool calls in one tutorial scene, use the batch endpoint:

curl -s -X POST http://127.0.0.1:32123/mcp/calls \
  -H 'Content-Type: application/json' \
  -d '{"calls":[{"tool":"clear","arguments":{}},{"tool":"openclicky_point","arguments":{"x":640,"y":520,"caption":"Start here"}},{"tool":"speak","arguments":{"text":"Start with this button."}}]}'

JSON-RPC style MCP calls are also accepted:

curl -s -X POST http://127.0.0.1:32123/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"openclicky_point","arguments":{"x":640,"y":520,"caption":"This one"}}}'

SSE status stream

For another app that wants acknowledgements:

curl -N http://127.0.0.1:32123/events

Behavior rules

When the user says "show me where you mean", "show me how to", "how do I do this", "point to it", or similar:

  1. Point only at a visible target that is directly relevant to the user's current question, instruction, or next step.
  2. If you already know that relevant coordinate, immediately call openclicky_point or /cursor with the best available coordinate.
  3. If you do not know the coordinate, immediately call /screenshot, inspect it, and point only if the relevant target is visible.
  4. If the relevant target is not visible or the connection is ambiguous, do not point; answer briefly or ask for the missing context instead.
  5. Keep captions short: 3-8 words is ideal.
  6. Do not start a new agent just to point.
  7. Do not narrate a long explanation first when a relevant target is visible; show the on-screen cue first, then add text only if needed.

Example response flow:

curl -s -X POST http://127.0.0.1:32123/cursor \
  -H 'Content-Type: application/json' \
  -d '{"x":1180,"y":760,"caption":"Use this menu", "durationMs":5000}'

Then reply briefly: "Shown on screen."

Scribble example:

curl -s -X POST http://127.0.0.1:32123/scribble \
  -H "Authorization: Bearer $OPENCLICKY_BRIDGE_TOKEN" \
  -H 'Content-Type: application/json' \
  -d '{"points":[{"x":640,"y":860},{"x":700,"y":900},{"x":760,"y":860}],"accentHex":"#F59E0B","lineWidth":6,"durationMs":3500,"caption":"Look here"}'

Rectangle highlight example:

curl -s -X POST http://127.0.0.1:32123/highlight \
  -H "Authorization: Bearer $OPENCLICKY_BRIDGE_TOKEN" \
  -H 'Content-Type: application/json' \
  -d '{"x":600,"y":720,"width":260,"height":140,"accentHex":"#60A5FA","fillOpacity":0.16,"durationMs":3500,"caption":"Important area"}'

Current visual tool boundary

The implemented local bridge is token-gated and currently exposes cursor pointing, multiple temporary cursor markers, captions, screenshot capture, coordinate click, clear, speak, notify, batch calls, and MCP tool descriptors.

The bridge now implements temporary freehand scribbles via /scribble / show_scribble and rectangle highlights via /highlight, /rectangle, show_highlight, and show_rectangle. Do not claim spotlight masks, arrows, area dimming, or persistent annotations exist until GET /mcp/tools lists those tools and OpenClickyExternalControlBridge.swift implements matching commands.

Before visual guidance, prefer GET /health or GET /mcp/tools when uncertain. Use only visible, current-screen targets and clear stale overlays with /clear.