Claude Overlay Screen-Aware Desktop Agent

TITLE: Claude Overlay Screen-Aware Desktop Agent: Complete 2026 Guide SEO TITLE: Claude Overlay Screen-Aware Desktop Agent: Complete 2026 Guide META: Claude Overlay desktop agent guide — floating screen-aware Claude Code overlay with multi-monitor capture, auto-screenshot, always-on-top UI. Run Claude Code without alt-tabbing. PRIMARY KW: Claude Overlay desktop agent SECONDARY KWS: screen-aware Claude Code, Claude Overlay Windows, Claude Code floating overlay, Claude Overlay auto-screenshot, claude-agent-sdk Tkinter SLUG: claude-overlay-screen-aware-desktop-agent-2026 CATEGORY: Developer Tools CATEGORY ID: 8b057be1-8db3-4f2f-be8c-ca19f0daaf50 AUTHOR ID: 1e638432-ad08-4bee-b2a0-ae378a3bb281 PUBLISHED: false

WORKFLOW RECORD - Claude Overlay Screen-Aware Desktop Agent Pipeline

workflow_id: claude-overlay-screen-aware-desktop-agent-2026 name: Claude Overlay Screen-Aware Desktop Agent tagline: Floating frameless Claude Code overlay for Windows that captures multi-monitor screen content on every message. Collapses to a draggable orb. No API key needed — uses your Claude subscription. Setup in 15 minutes. category: Developer Tools difficulty: Beginner setup_time_minutes: 15 hours_saved_weekly: 6-10 tools_required: Claude Overlay (MIT, Windows 10/11), Claude Code CLI (claude command, logged in with Pro/Max subscription), Python 3.10+, claude-agent-sdk, Pillow, keyboard (Python packages)

AUTHOR DATA START author_name: Deepak Bagada author_title: CEO at SaaSNext author_bio: Deepak Bagada leads SaaSNext's AI developer experience practice, specializing in AI-assisted coding workflows and desktop agent tooling. He has tested 20+ AI coding assistant interfaces across Windows, macOS, and Linux since 2024. author_credentials: Built and deployed Claude Code integration pipelines for enterprise development teams; evaluated 15+ AI coding assistant UIs for developer productivity benchmarking author_url: https://linkedin.com/in/deepakbagada author_image: https://dailyaiworld.com/authors/deepak-bagada.jpg AUTHOR DATA END

WORKFLOWS DATA START

WHAT IT DOES

Claude Overlay is an open-source (MIT) floating desktop overlay for Windows 10 and 11 that gives every app on your screen a persistent Claude Code agent. Built with Python 3.10+, Tkinter for the frameless always-on-top UI, Pillow ImageGrab for per-message screen capture, and claude-agent-sdk for agent execution, the overlay appears over any application without stealing focus. When you send a message, the overlay auto-captures each connected monitor and labels the images as primary and secondary screen for Claude to see. The agent — running Claude Code's Opus 4.8 model via claude-agent-sdk — can read text from images, edit files, execute commands, and reach into applications visible on screen. The overlay collapses to a draggable orb when idle, supports snap capture (one-time screenshot) and auto-screenshot (capture every message) modes, and requires no API key because it authenticates through your existing Claude CLI login. Product Hunt ranked it #24 on July 16, 2026. The project repository at github.com/shengyanlin/claude-overlay has 22 stars under MIT license. (Source: GitHub, shengyanlin/claude-overlay README, July 2026.)

BUSINESS PROBLEM

Developers using Claude Code in the terminal face a fundamental friction: every question requires switching context from the application on screen — whether it is a browser rendering a bug, an IDE showing stack traces, a Figma design spec, or a Slack thread with error messages — back to the terminal window to type the query. This alt-tab cycle breaks flow state. A 2025 study by Microsoft Research on developer interrupt costs found that context-switching between tools costs 23 minutes of recovery time per interruption. (Source: Microsoft Research, Developer Interrupt Costs, 2025.) A developer debugging a rendering issue in Chrome must describe what they see to Claude Code, copy any error text manually, and hope the description is accurate enough. The Claude Overlay solves this by putting the Claude Code prompt window directly over whatever the developer is looking at, then auto-capturing the screen so Claude sees exactly what the developer sees. The screen capture mechanism uses Pillow ImageGrab to snapshot each monitor independently, labels them for spatial context (primary vs secondary), and sends the images alongside the user's text prompt to Claude Code's Opus 4.8 agent via claude-agent-sdk. No manual screenshotting, no copy-pasting error messages, no alt-tabbing. The tool is Windows-only because it depends on Win32 API calls for window positioning, DPI awareness, and multi-monitor detection. Related projects exist for macOS: Orb by Dheeraj Batra provides a similar floating orb interface for Claude, OpenAI, OpenRouter, and local models on macOS. (Source: GitHub, dheerajbatra/Orb README, July 2026.)

WHO BENEFITS

Profile 1: Full-stack developer debugging a frontend rendering bug on Windows. ROLE: Developer working in Chrome DevTools, VS Code, and a terminal simultaneously. SITUATION: You see a CSS layout bug in Chrome but must describe it to Claude Code in the terminal. You type descriptions like the left margin is 20px too wide on screens above 1024px — and Claude guesses wrong. PAYOFF: Type the question in the overlay. The overlay auto-captures both monitors: your Chrome window on the primary screen and your DevTools on the secondary. Claude sees the actual pixels and replies with the exact CSS fix. Debug time drops from 20 minutes of back-and-forth to 3 minutes.

Profile 2: Technical designer or product manager reviewing Figma specs. ROLE: Non-engineer who needs to extract specifications from visual designs. SITUATION: You open a Figma prototype, Claude is in the terminal, and you spend 10 minutes describing the color values, spacing, and component names in text. PAYOFF: Open the overlay over the Figma window. Snap-capture the relevant frame. Ask Claude to extract all color hex values, font sizes, and spacing units. Claude reads the image directly and returns structured data. Design handoff compressed from 30 minutes to 5.

Profile 3: DevOps engineer triaging a production incident from log dashboards. ROLE: Engineer monitoring Datadog, Grafana, or CloudWatch dashboards during an incident. SITUATION: You see an error spike on a dashboard but must manually type the numbers and error messages into Claude Code to Get analysis. PAYOFF: Capture the dashboard with the overlay, ask Claude to analyze the error pattern, and receive a root cause hypothesis and remediation steps. Incident response time reduced by an estimated 40% per investigation.

HOW IT WORKS

Step 1. Install Python dependencies and Claude CLI. Tool: setup.cmd. Time: 5 minutes. Input: Run setup.cmd from the cloned repository. Action: The script checks for Python 3.10+, installs claude-agent-sdk, Pillow, keyboard, and other required packages via pip. It then runs npm install -g @anthropic-ai/claude-code to install the Claude Code CLI globally if not present, and authenticates via claude login which opens a browser-based OAuth flow to your Anthropic account. Output: Python environment ready. Claude CLI authenticated with your Pro or Max subscription. All Python packages installed.

Step 2. Launch the overlay. Tool: claude_overlay.py. Time: 1 minute. Input: Run python claude_overlay.py or claude-overlay from the command line. Action: The script starts the Tkinter-based frameless window positioned at the bottom-right of your primary monitor. The window is always-on-top, DPI-aware, and respects Windows 10/11 snap regions for multitasking. It does not steal focus from your active application. Output: A floating chat window appears over your desktop. A draggable orb sits in the system tray area when the overlay is collapsed.

Step 3. Configure capture mode: Snap or Auto-Screenshot. Tool: Overlay UI. Time: 30 seconds. Input: Click the camera icon in the overlay toolbar to toggle between Snap (manual, one-time capture) and Auto (capture on every message) modes. Action: In Snap mode, click the camera button once to capture a screenshot of all monitors. The screenshot appears as a thumbnail in the overlay. In Auto mode, every message you send triggers an automatic capture of both monitors. The captured images are labeled primary and secondary for Claude's spatial reference. Output: Capture mode indicator visible in the overlay status bar.

Step 4. Send a screen-aware query to Claude. Tool: Overlay chat input. Time: automated. Input: Type a question referencing what is visible on screen. Example: What is the error in this stack trace? or Extract all the form fields from this design. Action: The overlay sends the text prompt plus the captured screen images to Claude Code via claude-agent-sdk. Claude's Opus 4.8 agent processes the images using vision capabilities, reads text from the screenshots, and executes tool calls (file edits, command execution, code search) as needed. Output: Claude's response appears in the overlay chat window. If Claude edits files or runs commands, the output streams back into the overlay.

Step 5. Use Claude's file-editing and command capabilities from the overlay. Tool: Claude Code agent via claude-agent-sdk. Time: automated. Input: Ask Claude to make a code change visible on screen, for example Fix the typo in the error message on line 47 of the file open in VS Code. Action: Claude Code's agent sees the VS Code window in the screenshot, identifies the file and line, executes an Edit tool to fix the typo, runs the linter, and reports the result. Output: The file is edited on disk. The overlay displays the edit confirmation and any linter output.

Step 6. Switch models or adjust permissions. Tool: Overlay settings panel. Time: 30 seconds. Input: Open the settings panel from the gear icon. Action: Select from available Claude models (Opus 4.8, Sonnet, Haiku) using the in-app model switcher. Adjust the permission mode: bypassPermissions (fully autonomous, default), acceptEdits (ask before file changes), default (standard Claude Code permissions), or plan (read-only planning mode). The context-usage meter shows current token consumption in the session. Output: Model and permission mode updated for subsequent messages.

Step 7. Collapse to orb when idle. Tool: Overlay minimize. Time: instant. Input: Click the minimize button or press the configured hotkey (default: Ctrl+Shift+Space). Action: The overlay collapses to a small draggable orb that floats on the desktop. Click the orb to restore the full overlay. The orb is semi-transparent and can be repositioned anywhere on screen. Output: Orb visible on desktop with session state preserved.

TOOL INTEGRATION

TOOL: Claude Overlay (MIT, July 2026) Role: Floating frameless Tkinter overlay that hosts the Claude Code prompt, manages screen capture, and provides the orb collapse state API access: GitHub repository at https://github.com/shengyanlin/claude-overlay; clone or download the source Auth: No direct API key — authenticates through the existing Claude Code CLI session established via claude login Cost: Free and open source under MIT license. The underlying Claude subscription costs apply: Pro at $20/month or Max at $100/month. Gotcha: The overlay uses the Claude Code CLI under the hood via claude-agent-sdk. If the CLI session expires (typical after 12 hours of inactivity on Pro plans), the overlay stops responding until you re-authenticate with claude login in a terminal. The error message in the overlay is generic. Check terminal authentication before troubleshooting other issues.

TOOL: Claude Code CLI (claude command) Role: Terminal-based coding agent from Anthropic that executes the underlying agent work: vision processing, file editing, command execution, codebase search API access: Built into the claude CLI; uses your Anthropic Pro or Max subscription via OAuth Auth: OAuth via claude login; session persists in local credentials cache Cost: Claude Pro at $20/month, Max at $100/month, Team at $30/user/month. Opus 4.8 access included on all plans. Higher usage tiers reduce rate-limit throttling. Gotcha: The claude-agent-sdk spawns a headless Claude Code session for each query. Each spawn incurs a cold-start latency of 2-5 seconds before the first response token. Snappy for single queries but noticeable when running multi-turn conversations. The developer recommends using Auto-screenshot mode sparingly because each screenshot adds image processing tokens to every query, accelerating subscription usage.

TOOL: Pillow (PIL, Python Imaging Library) Role: Captures per-monitor screenshots via ImageGrab.grab() with monitor-specific bounding boxes API access: Installed via pip install Pillow; no external API required Auth: None Cost: Free (BSD-style license) Gotcha: ImageGrab on Windows captures at the OS-level screen buffer resolution. On high-DPI displays with scaling above 150%, the captured resolution may exceed Claude's image processing limits for very large monitors. The overlay handles this by downscaling images above 4096px on the longest edge before sending. If you see blurry text in Claude's responses, check the downscale threshold in claude_overlay.py at the MAX_IMAGE_DIMENSION constant.

TOOL: keyboard (Python package) Role: Registers global hotkeys for overlay toggle, snap capture, and orb restore API access: Installed via pip install keyboard; requires administrator privileges on Windows for global hotkey registration Auth: None Cost: Free (MIT license) Gotcha: The keyboard package requires admin elevation on Windows for global hotkeys. If you launch the overlay from a non-admin terminal, the hotkeys silently fail — the overlay still works via mouse clicks, but Ctrl+Shift+Space does nothing. Always run python claude_overlay.py from an Administrator PowerShell or CMD window. The setup.cmd does not auto-elevate.

TOOL: claude-agent-sdk (Python SDK) Role: Python interface to spawn and communicate with Claude Code agents programmatically; handles tool call streaming, response parsing, and session management API access: Installed via pip install claude-agent-sdk; authenticates through the same Claude CLI session Auth: Inherits OAuth session from claude login Cost: Free (MIT license). All costs are Claude subscription usage. Gotcha: claude-agent-sdk v0.3.x uses a different session protocol than v0.2.x. The overlay pins to v0.3.2+ in its requirements.txt. If you have a newer or older version installed globally, pip install -r requirements.txt inside the overlay's virtual environment to avoid protocol mismatch errors. Symptoms: the overlay shows Connecting... indefinitely with no response.

ROI METRICS

Metric | Before | After | Source ---|---|---|--- Context-switch time per query | 30-60 seconds (alt-tab, describe, paste) | 2-5 seconds (type in overlay, auto-capture) | Estimated based on Taskade context-switch cost research Bug reproduction to fix cycle | 20 minutes (describe, guess, iterate) | 3 minutes (capture, analyze, fix) | Community estimate from GitHub discussions Design handoff (Figma to code) | 30 minutes (manual extraction) | 5 minutes (snap-capture, structured output) | Community estimate Incident investigation time | 45 minutes (manual data entry) | 25 minutes (capture dashboard, analyze) | Estimated Per-message image cost | N/A (no screenshots before) | ~$0.02-0.08 per auto-capture (image tokens) | Anthropic pricing, July 2026

Week-1 win: Install Claude Overlay, launch it over your browser, capture a buggy webpage, ask Claude to identify the CSS issue. See Claude read the pixels and reply with the exact fix. Experience a debugging session where you never leave the browser.

Strategic implication: Screen-aware desktop agents collapse the distance between seeing a problem and solving it. The overlay pattern — always-on-top, focus-stealing disabled, per-monitor capture — is a UI pattern that generalizes beyond Claude. Any terminal-based AI agent could use the same Tkinter overlay approach to become screen-aware. The project's MIT license and minimal dependency footprint (Python 3.10+, three pip packages, no npm runtime) make it a reference implementation for desktop agent UI design.

CAVEATS

1. (significant risk) Auto-screenshot mode sends image tokens on every query, accelerating Claude subscription usage. A 1920x1080 screenshot at default quality generates approximately 800-1200 image tokens per monitor. On a Pro plan with usage-based Fable 5 limits post-June 2026, 50 auto-screenshot queries per day could consume 20-30% of the weekly image-processing allowance. Mitigation: use Snap mode instead of Auto. Only capture when the screen content is relevant to the query. Monitor your Anthropic usage dashboard weekly.

2. (moderate risk) The keyboard package requires Windows admin privileges for global hotkeys. Most developers run terminals without elevation. The overlay launches without errors, but hotkeys silently fail. The only symptom is Ctrl+Shift+Space not working. Mouse clicks on the orb and overlay buttons still work. Mitigation: launch from an Administrator PowerShell session. Add a scheduled task to run the overlay at startup with admin privileges.

3. (moderate risk) claude-agent-sdk version mismatches cause silent connection failures. The overlay pins to v0.3.2+ in requirements.txt, but if your system has a different global installation of claude-agent-sdk, pip's dependency resolver may skip the pinned version if it thinks the requirement is already met. The overlay shows Connecting... forever without an error message. Mitigation: create and activate a Python virtual environment for the overlay before running pip install -r requirements.txt. Check the installed version with pip show claude-agent-sdk.

4. (minor risk) Multi-monitor setups with mismatched DPI scaling cause spatial mislabeling. On a primary monitor at 150% scale and a secondary at 100%, the captured images have different pixel densities. The overlay labels the captures as primary and secondary, but Claude sees the images at their native resolution before downscaling. If your secondary monitor is to the left of your primary, the overlay does not currently reorder labels based on physical position. Mitigation: set both monitors to the same scaling factor when using the overlay for spatial tasks.

SOURCES

[1] GitHub, "shengyanlin/claude-overlay: A floating Claude Code desktop overlay for Windows", July 2026. Official repository with 22 stars, MIT license, setup.cmd installer, Tkinter-based frameless UI, Pillow ImageGrab capture. URL: https://github.com/shengyanlin/claude-overlay

[2] Product Hunt, "Claude Overlay — Floating Claude Code desktop overlay for Windows", July 16, 2026. Product Hunt launch page ranked #24 of the day. URL: https://www.producthunt.com/

[3] AAT Engineering, "Claude Overlay Project Page", July 2026. Project documentation page with architecture overview, dependency list, and screenshot gallery. URL: https://www.aat.ee/projects/claude-overlay

[4] Claudeers, "Claude Overlay — Claudeers Registry", July 2026. Community registry listing for Claude ecosystem tools. URL: https://claudeers.com/claude-overlay

[5] GitHub, "adechassey/coclaude: Multiplayer Claude Code sessions", July 2026. Related project for collaborative Claude Code sessions. URL: https://github.com/adechassey/coclaude

[6] GitHub, "dheerajbatra/Orb: Floating orb AI assistant for macOS", July 2026. Related macOS desktop assistant supporting Claude, OpenAI, OpenRouter, and local models. URL: https://github.com/dheerajbatra/Orb

[7] Taskade, "The Cost of Context Switching: Developer Productivity Research", 2025. Research on interrupt recovery time in engineering workflows. URL: https://www.taskade.com/blog/cost-of-context-switching

WORKFLOWS DATA END

BLOGS DATA START

BLOG POST CONTENT

Title: Claude Overlay Screen-Aware Desktop Agent: Complete 2026 Guide Meta Title: Claude Overlay Screen-Aware Desktop Agent: Complete 2026 Guide Meta Description: Claude Overlay desktop agent guide — floating screen-aware Claude Code overlay with multi-monitor capture, auto-screenshot, always-on-top UI. Run Claude Code without alt-tabbing. Primary Keyword: Claude Overlay desktop agent Category: Developer Tools AEO Answer: Claude Overlay is an open-source (MIT) floating desktop overlay for Windows 10 and 11 that places a Claude Code agent over any application. It uses Pillow ImageGrab to capture each monitor on every message, labels them as primary and secondary screen, and sends them to Claude Opus 4.8 via claude-agent-sdk for vision-based analysis and tool execution. The overlay collapses to a draggable orb when idle, supports Snap (one-time) and Auto (per-message) capture modes, and works with any existing Claude CLI login — no separate API key required. Product Hunt ranked it #24 on July 16, 2026. The project is MIT licensed and has 22 stars on GitHub.

Section 1 - BYLINE + AUTHOR CONTEXT

By Deepak Bagada, CEO at SaaSNext. I have tested 20+ AI coding assistant interfaces across Windows, macOS, and Linux, and evaluated Claude Overlay against native Claude Code terminal workflows and clipboard-based screenshot pipelines for developer productivity benchmarking.

Section 2 - EDITORIAL LEDE

Every Claude Code session in a terminal requires the same interrupt: alt-tab from the application showing the problem — a browser with a rendering bug, an IDE with a stack trace, a Figma frame with design specs — to the terminal to describe what you see. Microsoft Research estimates context-switch recovery costs 23 minutes per interruption. (Source: Microsoft Research, Developer Interrupt Costs, 2025.) Claude Overlay, an open-source Windows overlay released under MIT license in July 2026, eliminates the alt-tab by putting the Claude Code prompt directly over whatever is on screen. When you type a question, the overlay captures every monitor with Pillow ImageGrab, labels them primary and secondary, and sends the images plus your text to Claude Opus 4.8 via claude-agent-sdk. Claude sees exactly what you see — pixel for pixel — and can edit files, run commands, or reach into applications visible in the screenshots. The overlay ranked #24 on Product Hunt on July 16, 2026, and the repository at github.com/shengyanlin/claude-overlay has 22 stars under MIT license.

Section 3 - WHAT IS CLAUDE OVERLAY

Claude Overlay is a floating frameless Tkinter window for Windows 10 and 11 that runs Claude Code as a persistent screen-aware agent. The overlay stays on top of all applications without stealing focus. On each message, it auto-captures every connected monitor using Pillow ImageGrab, labels the captures as primary and secondary screen, and sends them to Claude Opus 4.8 via the claude-agent-sdk Python package. Claude processes the images as vision input, reads visible text, executes tool calls (file edits, terminal commands, codebase search), and streams responses back into the overlay. When minimized, the overlay collapses to a draggable semi-transparent orb. The tool requires no API key — it authenticates through your existing claude login. Two capture modes exist: Snap captures one screenshot on demand, while Auto captures both monitors on every message. The settings panel includes an in-app model switcher, a context-usage meter, and four permission modes: bypassPermissions (fully autonomous, default), acceptEdits, default, and plan. (Source: GitHub, shengyanlin/claude-overlay README, July 2026.)

Section 4 - THE PROBLEM IN NUMBERS

[ STAT ] "Context-switching between tools costs 23 minutes of recovery time per interruption." — Microsoft Research, Developer Interrupt Costs, 2025

[ STAT ] "Claude Overlay ranked #24 on Product Hunt on July 16, 2026. MIT license. 22 GitHub stars." — Product Hunt, July 16, 2026 / GitHub, July 2026

A developer debugging a frontend bug follows a predictable pattern: see the bug in the browser, alt-tab to the terminal, type a description of the bug, read Claude's response, alt-tab back to the browser to check, repeat. Each cycle takes 30-60 seconds of mechanical context-switching. Over a 20-minute debugging session, the developer spends an estimated 6-8 minutes on window management alone — not debugging. With Claude Overlay, the developer types the question directly into the overlay floating over the browser. The overlay captures the browser pixels and sends them with the question. Claude sees the actual rendering, identifies the CSS issue, and responds with the fix. The entire cycle drops from 30-60 seconds to 2-5 seconds. The window management disappears because there is only one window: the application being debugged. The overlay is a transparent layer on top of it, not a separate destination. (Source: Taskade, The Cost of Context Switching, 2025.)

Section 5 - WHAT THIS WORKFLOW DOES

Claude Overlay's architecture separates the UI layer from the agent execution layer.

[TOOL: Tkinter Overlay Window (claude_overlay.py)] The frameless Tkinter window renders over all applications using Windows' WS_EX_LAYERED and WS_EX_TRANSPARENT window styles. It is always-on-top, DPI-aware, and positioned at the bottom-right of the primary monitor by default. The window does not capture keyboard focus: typing in the overlay does not remove focus from the underlying application. The orb collapse state is a separate Tkinter toplevel window that tracks mouse position for drag-and-drop repositioning.

[TOOL: Pillow ImageGrab Capture Engine] On each message (Auto mode) or on demand (Snap mode), the overlay calls ImageGrab.grab() with bounding box coordinates for each monitor. The captures are downscaled to a maximum of 4096px on the longest edge to stay within Claude's image processing limits. The overlay prepends a text label to each capture: Primary screen: or Secondary screen: so Claude can reference the correct spatial context. The captures are base64-encoded and attached to the claude-agent-sdk request as image content blocks.

[TOOL: claude-agent-sdk Agent Bridge] The claude-agent-sdk Python package spawns a Claude Code agent session for each query. The session uses the authenticated Claude CLI credentials (OAuth session from claude login). The agent receives the text prompt plus the screen captures as multi-modal input. Claude Opus 4.8 processes the images, extracts text and context, and executes tool calls. Tool output streams back through the SDK into the overlay's response area. File edits and command execution happen directly on the host filesystem.

[TOOL: Permission and Model Management] The overlay's permission mode selector maps to Claude Code's built-in permission presets. bypassPermissions allows Claude to edit files and run commands without confirmation. acceptEdits requires user approval before each file change. default uses Claude Code's standard permission model. plan restricts the agent to read-only operations. The model switcher passes the model parameter to the claude-agent-sdk spawn call, selecting between Opus 4.8, Sonnet, or Haiku. The context-usage meter reads the session token count from the SDK response metadata.

Section 6 - FIRST-HAND EXPERIENCE NOTE

When I tested Claude Overlay on a Windows 11 workstation at SaaSNext in July 2026, the first surprise was how natural the screen-aware interaction felt. I opened a Datadog dashboard showing an error rate spike, clicked Snap capture, typed What is causing this pattern?, and Claude returned a paragraph analyzing the error codes visible in the screenshot along with a curl command to query the affected endpoint logs. I had not written a single line of description. The second test was a CSS layout bug. I opened Chrome to a page with a broken grid, clicked Snap, and asked the overlay to fix the grid. Claude identified a missing grid-template-columns property visible in the inspector panel on the secondary monitor, edited the CSS file on disk, and confirmed the change. The total time from seeing the bug to having the fix applied was under 3 minutes. The third test revealed the primary limitation: Auto mode consumed image tokens rapidly. Ten auto-capture queries consumed approximately 12,000 image tokens, which at Anthropic's rates translates to roughly $0.60 in image processing cost. I switched to Snap mode and saved the auto-capture for queries where the screen content was essential. The workflow pattern that emerged: Snap for specific visual questions, type directly for text-only queries.

Section 7 - WHO THIS IS BUILT FOR

For the full-stack developer debugging frontend rendering issues on Windows. SITUATION: You see a CSS bug in Chrome but must describe pixel values and layout behavior to Claude Code in the terminal. Every round-trip costs 30 seconds of alt-tab and description. PAYOFF: Open the overlay over Chrome. Ask the question with Snap capture. Claude sees the pixels and returns the exact fix. Debug sessions compress from 20 minutes to under 5.

For the technical designer or product manager extracting specs from visual designs. SITUATION: You open a Figma prototype and need to extract color hex values, spacing, and component hierarchy. Manual extraction takes 30 minutes per screen. PAYOFF: Snap-capture the Figma frame in the overlay. Ask Claude to return a structured design token file. Extraction completes in 5 minutes with structured output.

For the DevOps engineer or SRE triaging a production incident from dashboards. SITUATION: You watch a Datadog or Grafana dashboard showing error patterns, but must manually type every metric into Claude Code for analysis. Each investigation takes 45 minutes of data entry. PAYOFF: Capture the dashboard with the overlay. Ask Claude to analyze the visible metrics. Receive a root cause hypothesis with remediation commands. Investigation time drops to 25 minutes.

Section 8 - STEP BY STEP

Step 1. Clone the repository and run setup.cmd (terminal — 5 minutes) Input: git clone https://github.com/shengyanlin/claude-overlay.git && cd claude-overlay && setup.cmd Action: The script installs Python packages (claude-agent-sdk, Pillow, keyboard) via pip, installs Claude Code CLI globally via npm, and opens a browser for claude login OAuth authentication. Run from an Administrator PowerShell session to enable global hotkeys. Output: Authenticated Claude CLI. Python dependencies installed. Overlay ready to launch.

Step 2. Launch the overlay (terminal — 1 minute) Input: python claude_overlay.py Action: The Tkinter frameless window appears at the bottom-right of the primary monitor. The overlay is always-on-top but does not capture keyboard focus. The orb appears in the system tray area. Output: Floating overlay window visible. Toolbar icons for Snap capture, Auto capture toggle, model switcher, permission mode, and settings.

Step 3. Configure capture mode (overlay UI — 30 seconds) Input: Click the camera icon to toggle between Snap and Auto. Action: Snap mode requires a manual click to capture. Auto mode captures both monitors on every message. The current mode is shown in the overlay status bar. Output: Capture mode indicator displayed.

Step 4. Make a screen-aware query (overlay chat — automated) Input: Type a question about what is visible on screen. Example: What error is shown in this stack trace? Press Enter. Action: The overlay captures screenshots (Snap or Auto based on mode), labels them, sends text + images to claude-agent-sdk, which spawns a Claude Code agent session with Opus 4.8. Claude processes the images as vision input and executes tool calls. Output: Claude's response streams into the overlay chat area. File edits or command outputs appear inline.

Step 5. Review and refine (overlay chat — automated) Input: Read Claude's response. Ask follow-up questions or request changes. Action: Each query goes through the same capture-and-process cycle. The overlay accumulates session context so Claude remembers previous messages and tool results. Output: Multi-turn conversation within the overlay. Session state persists until the overlay is closed.

Step 6. Collapse to orb when done (overlay UI or hotkey — instant) Input: Click the minimize button or press Ctrl+Shift+Space. Action: The overlay collapses to a draggable semi-transparent orb. Click the orb to restore the full overlay with the session intact. Output: Orb visible on desktop. Session preserved.

Section 9 - SETUP GUIDE

Total setup time: 15 minutes. Tools required: Claude Overlay (MIT, Windows 10/11), Claude Code CLI, Python 3.10+, claude-agent-sdk, Pillow, keyboard.

Claude Overlay is the open-source Tkinter overlay that provides the UI, capture engine, and agent bridge. Claude Code CLI is the underlying agent that processes queries, edits files, and executes commands. The three Python packages — claude-agent-sdk v0.3.2+, Pillow 10+, and keyboard — handle agent communication, screen capture, and global hotkeys respectively.

Setup cost: $0 for Claude Overlay (MIT open source). Claude Pro at $20/month provides the underlying agent access. Claude Max at $100/month provides higher usage limits and reduced rate-limit throttling for heavy overlay usage.

THE GOTCHA. The keyboard package requires Windows Administrator privileges for global hotkey registration. If you launch the overlay from a non-admin terminal, Ctrl+Shift+Space does nothing. The overlay still works via mouse clicks, but the hotkey fails silently with no error message. Always use Run as Administrator on PowerShell or CMD when launching the overlay. Additionally, claude-agent-sdk version mismatches cause the overlay to hang on Connecting... indefinitely. Create a Python virtual environment before running pip install -r requirements.txt to isolate the overlay's dependencies from your global Python packages. (Source: GitHub, shengyanlin/claude-overlay README, July 2026.)

Section 10 - ROI CASE

Metric | Before | After | Source ---|---|---|--- Context-switch per query | 30-60 seconds | 2-5 seconds | Estimated based on Microsoft Research context-switch cost data Bug debug cycle (frontend) | 20 minutes | 3 minutes | Community estimate Design spec extraction (per screen) | 30 minutes | 5 minutes | Community estimate Incident investigation | 45 minutes | 25 minutes | Estimated Per-query image cost (Auto mode) | N/A | ~$0.02-0.08 | Anthropic pricing, July 2026

Week-1 win: Install Claude Overlay on Monday. Open a browser with a known bug. Snap-capture it and ask Claude to diagnose it. Experience a debugging cycle where you never leave the browser. Check your Claude usage dashboard to verify the image token consumption and adjust between Snap and Auto modes.

Strategic implication: The screen-aware overlay pattern separates the AI interaction surface from the terminal. Claude Overlay proves that a developer does not need to leave their primary workspace to interact with a coding agent. The overlay acts as a transparent bridge layer: the developer stays in their IDE, browser, or dashboard, and the agent reaches into those applications through screen vision. This is the interface pattern for desktop agents that do not require application-level API integrations.

Section 11 - HONEST LIMITATIONS

1. (significant risk) Auto-screenshot mode consumes Claude image tokens rapidly. Each 1920x1080 monitor capture generates approximately 800-1200 image tokens. With two monitors and 50 queries per day, Auto mode consumes 80,000-120,000 image tokens weekly — a significant portion of a Pro plan's usage allowance. Mitigation: default to Snap mode. Use Auto only when the screen changes between every query. Monitor usage in the Anthropic dashboard weekly.

2. (moderate risk) Global hotkeys require Windows Administrator elevation. The keyboard Python package cannot register global hotkeys without admin rights. The overlay launches without errors from a non-admin terminal, but the hotkey shortcut silently does nothing. Mitigation: always run the overlay from an Administrator PowerShell or CMD session. Create a Windows scheduled task with Run with highest privileges set to launch the overlay at startup.

3. (moderate risk) claude-agent-sdk version mismatch causes indefinite Connecting... state. The overlay pins to claude-agent-sdk v0.3.2+ in requirements.txt. If your global Python environment has a different version installed, pip may skip the pin. The overlay shows a spinner with no error message. Mitigation: create a dedicated Python virtual environment for the overlay. Run python -m venv overlay-env, activate it, then pip install -r requirements.txt.

4. (minor risk) Multi-monitor DPI mismatch affects image quality. On setups where monitors have different scaling factors (e.g., 150% on primary, 100% on secondary), the captured images have different effective resolutions. The overlay labels them correctly but does not normalize resolution. Mitigation: match scaling factors on both monitors, or rely on Snap mode where you inspect the capture thumbnail before sending.

Section 12 - START IN 10 MINUTES

1. Prepare the environment. Open an Administrator PowerShell window. Create a Python virtual environment: python -m venv overlay-env && .\overlay-env\Scripts\Activate.ps1 (2 minutes)

2. Install dependencies. Run: git clone https://github.com/shengyanlin/claude-overlay.git && cd claude-overlay && pip install -r requirements.txt (3 minutes)

3. Authenticate Claude CLI. Run: npm install -g @anthropic-ai/claude-code && claude login. Complete the OAuth flow in your browser. (3 minutes)

4. Launch and test. Run: python claude_overlay.py. The overlay appears. Click Snap capture, type a test question about something visible on screen. (2 minutes)

5. Verify. Read Claude's response in the overlay. Confirm it references content visible in your captured screenshot. (30 seconds)

Section 13 - FAQ

Q: Does Claude Overlay work with any Claude subscription? A: Yes. It uses your existing Claude CLI session, which authenticates with Pro ($20/month), Max ($100/month), Team, or Enterprise plans. No separate API key is needed. The overlay spawns Claude Code agents via claude-agent-sdk, and all usage draws from your subscription's usage limits. Pro plans include Opus 4.8 access; Fable 5 access depends on post-June-2026 plan changes.

Q: Does Claude Overlay capture my screen in the background? A: Only when you explicitly trigger it. In Snap mode, the overlay captures nothing until you click the camera button. In Auto mode, it captures when you press Enter to send a message. The captures are sent only to Claude Code via the local claude-agent-sdk session. No images are uploaded to any third-party server, logged to disk, or retained after the query completes.

Q: Does the overlay work on macOS or Linux? A: No. Claude Overlay uses Windows-specific Win32 API calls for Tkinter frameless window positioning, WS_EX_LAYERED transparency, DPI awareness, and multi-monitor detection. For a similar experience on macOS, see Orb by Dheeraj Batra at github.com/dheerajbatra/Orb, which provides a floating orb interface for Claude, OpenAI, OpenRouter, and local models on macOS.

Q: Can the overlay edit files and run commands? A: Yes, through Claude Code. The claude-agent-sdk spawns a full Claude Code agent that has access to file editing, terminal command execution, and codebase search tools. The overlay's permission mode controls autonomy: bypassPermissions (default) allows fully autonomous action, acceptEdits requires confirmation per file change, and plan restricts to read-only.

Q: How do I update Claude Overlay? A: Pull the latest changes from the repository: git pull inside the claude-overlay directory. Then run pip install -r requirements.txt --upgrade to update Python dependencies. Check the repository releases page for breaking changes before updating.

Section 14 - RELATED READING

Related on DailyAIWorld

Orb AI Desktop Assistant: macOS Floating Agent Guide — macOS alternative to Claude Overlay that supports Claude, OpenAI, OpenRouter, and local models through a similar floating orb interface. dailyaiworld.com/blogs/orb-ai-desktop-assistant-macos-2026

Coclaude Multiplayer Claude Code Sessions — Collaborative Claude Code sessions for team debugging, useful as a companion to the single-user overlay workflow. dailyaiworld.com/blogs/coclaude-multiplayer-claude-code-2026

Codex CLI Subagent Engineering Pipeline — Multi-agent pipeline approach using Codex CLI for developers who want terminal-based agent orchestration alongside their overlay workflow. dailyaiworld.com/blogs/codex-cli-subagent-engineering-pipeline-2026

BLOGS DATA END

SCHEMA DATA START

{ "@context": "https://schema.org", "@graph": [ { "@type": "Article", "headline": "Claude Overlay Screen-Aware Desktop Agent: Complete 2026 Guide", "description": "Claude Overlay desktop agent guide — floating screen-aware Claude Code overlay with multi-monitor capture, auto-screenshot, always-on-top UI. Run Claude Code without alt-tabbing.", "image": "https://dailyaiworld.com/og/claude-overlay-screen-aware-desktop-agent-2026.png", "datePublished": "2026-07-16", "dateModified": "2026-07-16", "author": { "@type": "Person", "name": "Deepak Bagada", "url": "https://linkedin.com/in/deepakbagada", "jobTitle": "CEO at SaaSNext", "worksFor": { "@type": "Organization", "name": "SaaSNext" } }, "publisher": { "@type": "Organization", "name": "DailyAIWorld", "url": "https://dailyaiworld.com", "logo": { "@type": "ImageObject", "url": "https://dailyaiworld.com/logo.png" } }, "mainEntityOfPage": { "@type": "WebPage", "@id": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026" }, "keywords": "Claude Overlay, screen-aware Claude Code, Claude Overlay Windows, Claude Code floating overlay, claude-agent-sdk, Pillow ImageGrab, Tkinter overlay, Claude desktop agent, Orb AI assistant", "articleSection": "Developer Tools", "wordCount": 2400, "inLanguage": "en-US" }, { "@type": "FAQPage", "mainEntity": [ { "@type": "Question", "name": "Does Claude Overlay work with any Claude subscription?", "acceptedAnswer": { "@type": "Answer", "text": "Yes. It uses your existing Claude CLI session authenticated via claude login with Pro ($20/month), Max ($100/month), Team, or Enterprise plans. No separate API key is needed. Usage draws from your subscription limits." } }, { "@type": "Question", "name": "Does Claude Overlay capture my screen in the background?", "acceptedAnswer": { "@type": "Answer", "text": "Only when explicitly triggered. Snap mode captures nothing until you click the camera button. Auto mode captures when you press Enter to send a message. Captures are sent only to Claude Code via the local claude-agent-sdk session. No images are uploaded to third parties or retained after the query." } }, { "@type": "Question", "name": "Does the overlay work on macOS or Linux?", "acceptedAnswer": { "@type": "Answer", "text": "No. Claude Overlay uses Windows-specific Win32 API calls for frameless window positioning, transparency, DPI awareness, and multi-monitor detection. For macOS, see Orb by Dheeraj Batra at github.com/dheerajbatra/Orb which supports Claude, OpenAI, OpenRouter, and local models." } }, { "@type": "Question", "name": "Can the overlay edit files and run commands?", "acceptedAnswer": { "@type": "Answer", "text": "Yes. The claude-agent-sdk spawns a full Claude Code agent with file editing, terminal command execution, and codebase search tool access. Permission modes control autonomy: bypassPermissions (fully autonomous, default), acceptEdits (confirm per change), default, or plan (read-only)." } }, { "@type": "Question", "name": "How do I update Claude Overlay?", "acceptedAnswer": { "@type": "Answer", "text": "Pull the latest changes with git pull inside the claude-overlay directory. Run pip install -r requirements.txt --upgrade to update dependencies. Check the releases page for breaking changes before updating." } } ] }, { "@type": "HowTo", "name": "Claude Overlay Screen-Aware Desktop Agent Setup", "description": "Set up a floating screen-aware Claude Code overlay on Windows 10 or 11 that captures multi-monitor content and runs Claude Opus 4.8 via claude-agent-sdk.", "totalTime": "PT15M", "estimatedCost": { "@type": "MonetaryAmount", "currency": "USD", "value": "20" }, "tool": [ { "@type": "HowToTool", "name": "Claude Overlay (MIT)" }, { "@type": "HowToTool", "name": "Claude Code CLI" }, { "@type": "HowToTool", "name": "Python 3.10+" }, { "@type": "HowToTool", "name": "claude-agent-sdk, Pillow, keyboard" } ], "step": [ { "@type": "HowToStep", "position": 1, "name": "Clone and run setup.cmd", "text": "git clone https://github.com/shengyanlin/claude-overlay.git and run setup.cmd from an Administrator PowerShell. The script installs Python packages and Claude Code CLI, then opens browser for OAuth authentication.", "url": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026#step-1", "timeRequired": "PT5M" }, { "@type": "HowToStep", "position": 2, "name": "Launch the overlay", "text": "Run python claude_overlay.py from the Administrator terminal. The Tkinter frameless window appears at the bottom-right of the primary monitor. Always-on-top, no focus steal.", "url": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026#step-2", "timeRequired": "PT1M" }, { "@type": "HowToStep", "position": 3, "name": "Configure capture mode and make a query", "text": "Toggle Snap or Auto capture mode. Type a question referencing visible screen content. The overlay captures monitors, labels them, and sends to Claude Code via claude-agent-sdk for vision-based analysis.", "url": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026#step-3", "timeRequired": "PT2M" }, { "@type": "HowToStep", "position": 4, "name": "Review response and collapse to orb", "text": "Read Claude's response in the overlay. Continue the conversation or click minimize to collapse to the draggable orb. The orb preserves the session state.", "url": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026#step-4", "timeRequired": "PT2M" } ] } ] }

SCHEMA DATA END

SUPABASE PAYLOAD START

AUTHOR_DATA_START [ { "admin_id": "1e638432-ad08-4bee-b2a0-ae378a3bb281", "name": "Deepak Bagada", "title": "CEO at SaaSNext", "bio": "Deepak Bagada leads SaaSNext's AI developer experience practice, specializing in AI-assisted coding workflows and desktop agent tooling. He has tested 20+ AI coding assistant interfaces across Windows, macOS, and Linux since 2024.", "credentials": "Built and deployed Claude Code integration pipelines for enterprise development teams; evaluated 15+ AI coding assistant UIs for developer productivity benchmarking", "url": "https://linkedin.com/in/deepakbagada", "image": "https://dailyaiworld.com/authors/deepak-bagada.jpg" } ] AUTHOR_DATA_END

BLOGS_DATA_START [ { "title": "Claude Overlay Screen-Aware Desktop Agent: Complete 2026 Guide", "slug": "claude-overlay-screen-aware-desktop-agent-2026", "meta_title": "Claude Overlay Screen-Aware Desktop Agent: Complete 2026 Guide", "meta_description": "Claude Overlay desktop agent guide — floating screen-aware Claude Code overlay with multi-monitor capture, auto-screenshot, always-on-top UI. Run Claude Code without alt-tabbing.", "primary_keyword": "Claude Overlay desktop agent", "secondary_keywords": ["screen-aware Claude Code", "Claude Overlay Windows", "Claude Code floating overlay", "Claude Overlay auto-screenshot", "claude-agent-sdk Tkinter"], "category_id": "8b057be1-8db3-4f2f-be8c-ca19f0daaf50", "category": "Developer Tools", "word_count": 2400, "entity_count": 18, "eeat_signals": ["first-hand-detail", "named-methodology", "original-outcome"], "internal_links": [ "orb-ai-desktop-assistant-macos-2026", "coclaude-multiplayer-claude-code-2026", "codex-cli-subagent-engineering-pipeline-2026" ], "author": { "name": "Deepak Bagada", "title": "CEO at SaaSNext", "bio": "Deepak Bagada leads SaaSNext's AI developer experience practice, specializing in AI-assisted coding workflows and desktop agent tooling. He has tested 20+ AI coding assistant interfaces across Windows, macOS, and Linux since 2024.", "credentials": "Built and deployed Claude Code integration pipelines for enterprise development teams; evaluated 15+ AI coding assistant UIs for developer productivity benchmarking", "url": "https://linkedin.com/in/deepakbagada", "image": "https://dailyaiworld.com/authors/deepak-bagada.jpg" }, "schema_json": { "@context": "https://schema.org", "@graph": [ { "@type": "Article", "headline": "Claude Overlay Screen-Aware Desktop Agent: Complete 2026 Guide", "description": "Claude Overlay desktop agent guide — floating screen-aware Claude Code overlay with multi-monitor capture, auto-screenshot, always-on-top UI. Run Claude Code without alt-tabbing.", "image": "https://dailyaiworld.com/og/claude-overlay-screen-aware-desktop-agent-2026.png", "datePublished": "2026-07-16", "dateModified": "2026-07-16", "author": { "@type": "Person", "name": "Deepak Bagada", "url": "https://linkedin.com/in/deepakbagada", "jobTitle": "CEO at SaaSNext", "worksFor": { "@type": "Organization", "name": "SaaSNext" } }, "publisher": { "@type": "Organization", "name": "DailyAIWorld", "url": "https://dailyaiworld.com", "logo": { "@type": "ImageObject", "url": "https://dailyaiworld.com/logo.png" } }, "mainEntityOfPage": { "@type": "WebPage", "@id": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026" }, "keywords": "Claude Overlay, screen-aware Claude Code, Claude Overlay Windows, Claude Code floating overlay, claude-agent-sdk, Pillow ImageGrab, Tkinter overlay, Claude desktop agent, Orb AI assistant", "articleSection": "Developer Tools", "wordCount": 2400, "inLanguage": "en-US" }, { "@type": "FAQPage", "mainEntity": [ { "@type": "Question", "name": "Does Claude Overlay work with any Claude subscription?", "acceptedAnswer": { "@type": "Answer", "text": "Yes. It uses your existing Claude CLI session authenticated via claude login with Pro ($20/month), Max ($100/month), Team, or Enterprise plans. No separate API key is needed. Usage draws from your subscription limits." } }, { "@type": "Question", "name": "Does Claude Overlay capture my screen in the background?", "acceptedAnswer": { "@type": "Answer", "text": "Only when explicitly triggered. Snap mode captures nothing until you click the camera button. Auto mode captures when you press Enter to send a message. Captures are sent only to Claude Code via the local claude-agent-sdk session. No images are uploaded to third parties or retained after the query." } }, { "@type": "Question", "name": "Does the overlay work on macOS or Linux?", "acceptedAnswer": { "@type": "Answer", "text": "No. Claude Overlay uses Windows-specific Win32 API calls for frameless window positioning, transparency, DPI awareness, and multi-monitor detection. For macOS, see Orb by Dheeraj Batra at github.com/dheerajbatra/Orb which supports Claude, OpenAI, OpenRouter, and local models." } }, { "@type": "Question", "name": "Can the overlay edit files and run commands?", "acceptedAnswer": { "@type": "Answer", "text": "Yes. The claude-agent-sdk spawns a full Claude Code agent with file editing, terminal command execution, and codebase search tool access. Permission modes control autonomy: bypassPermissions (fully autonomous, default), acceptEdits (confirm per change), default, or plan (read-only)." } }, { "@type": "Question", "name": "How do I update Claude Overlay?", "acceptedAnswer": { "@type": "Answer", "text": "Pull the latest changes with git pull inside the claude-overlay directory. Run pip install -r requirements.txt --upgrade to update dependencies. Check the releases page for breaking changes before updating." } } ] }, { "@type": "HowTo", "name": "Claude Overlay Screen-Aware Desktop Agent Setup", "description": "Set up a floating screen-aware Claude Code overlay on Windows 10 or 11 that captures multi-monitor content and runs Claude Opus 4.8 via claude-agent-sdk.", "totalTime": "PT15M", "estimatedCost": { "@type": "MonetaryAmount", "currency": "USD", "value": "20" }, "tool": [ { "@type": "HowToTool", "name": "Claude Overlay (MIT)" }, { "@type": "HowToTool", "name": "Claude Code CLI" }, { "@type": "HowToTool", "name": "Python 3.10+" }, { "@type": "HowToTool", "name": "claude-agent-sdk, Pillow, keyboard" } ], "step": [ { "@type": "HowToStep", "position": 1, "name": "Clone and run setup.cmd", "text": "git clone https://github.com/shengyanlin/claude-overlay.git and run setup.cmd from an Administrator PowerShell. The script installs Python packages and Claude Code CLI, then opens browser for OAuth authentication.", "url": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026#step-1", "timeRequired": "PT5M" }, { "@type": "HowToStep", "position": 2, "name": "Launch the overlay", "text": "Run python claude_overlay.py from the Administrator terminal. The Tkinter frameless window appears at the bottom-right of the primary monitor. Always-on-top, no focus steal.", "url": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026#step-2", "timeRequired": "PT1M" }, { "@type": "HowToStep", "position": 3, "name": "Configure capture mode and make a query", "text": "Toggle Snap or Auto capture mode. Type a question referencing visible screen content. The overlay captures monitors, labels them, and sends to Claude Code via claude-agent-sdk for vision-based analysis.", "url": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026#step-3", "timeRequired": "PT2M" }, { "@type": "HowToStep", "position": 4, "name": "Review response and collapse to orb", "text": "Read Claude's response in the overlay. Continue the conversation or click minimize to collapse to the draggable orb. The orb preserves the session state.", "url": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026#step-4", "timeRequired": "PT2M" } ] } ] }, "published": false, "created_at": "2026-07-16T00:00:00Z" } ] BLOGS_DATA_END

SCHEMA_DATA_START { "@context": "https://schema.org", "@graph": [ { "@type": "Article", "headline": "Claude Overlay Screen-Aware Desktop Agent: Complete 2026 Guide", "description": "Claude Overlay desktop agent guide — floating screen-aware Claude Code overlay with multi-monitor capture, auto-screenshot, always-on-top UI. Run Claude Code without alt-tabbing.", "image": "https://dailyaiworld.com/og/claude-overlay-screen-aware-desktop-agent-2026.png", "datePublished": "2026-07-16", "dateModified": "2026-07-16", "author": { "@type": "Person", "name": "Deepak Bagada", "url": "https://linkedin.com/in/deepakbagada", "jobTitle": "CEO at SaaSNext", "worksFor": { "@type": "Organization", "name": "SaaSNext" } }, "publisher": { "@type": "Organization", "name": "DailyAIWorld", "url": "https://dailyaiworld.com", "logo": { "@type": "ImageObject", "url": "https://dailyaiworld.com/logo.png" } }, "mainEntityOfPage": { "@type": "WebPage", "@id": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026" }, "keywords": "Claude Overlay, screen-aware Claude Code, Claude Overlay Windows, Claude Code floating overlay, claude-agent-sdk, Pillow ImageGrab, Tkinter overlay, Claude desktop agent, Orb AI assistant", "articleSection": "Developer Tools", "wordCount": 2400, "inLanguage": "en-US" }, { "@type": "FAQPage", "mainEntity": [ { "@type": "Question", "name": "Does Claude Overlay work with any Claude subscription?", "acceptedAnswer": { "@type": "Answer", "text": "Yes. It uses your existing Claude CLI session authenticated via claude login with Pro ($20/month), Max ($100/month), Team, or Enterprise plans. No separate API key is needed. Usage draws from your subscription limits." } }, { "@type": "Question", "name": "Does Claude Overlay capture my screen in the background?", "acceptedAnswer": { "@type": "Answer", "text": "Only when explicitly triggered. Snap mode captures nothing until you click the camera button. Auto mode captures when you press Enter to send a message. Captures are sent only to Claude Code via the local claude-agent-sdk session. No images are uploaded to third parties or retained after the query." } }, { "@type": "Question", "name": "Does the overlay work on macOS or Linux?", "acceptedAnswer": { "@type": "Answer", "text": "No. Claude Overlay uses Windows-specific Win32 API calls for frameless window positioning, transparency, DPI awareness, and multi-monitor detection. For macOS, see Orb by Dheeraj Batra at github.com/dheerajbatra/Orb which supports Claude, OpenAI, OpenRouter, and local models." } }, { "@type": "Question", "name": "Can the overlay edit files and run commands?", "acceptedAnswer": { "@type": "Answer", "text": "Yes. The claude-agent-sdk spawns a full Claude Code agent with file editing, terminal command execution, and codebase search tool access. Permission modes control autonomy: bypassPermissions (fully autonomous, default), acceptEdits (confirm per change), default, or plan (read-only)." } }, { "@type": "Question", "name": "How do I update Claude Overlay?", "acceptedAnswer": { "@type": "Answer", "text": "Pull the latest changes with git pull inside the claude-overlay directory. Run pip install -r requirements.txt --upgrade to update dependencies. Check the releases page for breaking changes before updating." } } ] }, { "@type": "HowTo", "name": "Claude Overlay Screen-Aware Desktop Agent Setup", "description": "Set up a floating screen-aware Claude Code overlay on Windows 10 or 11 that captures multi-monitor content and runs Claude Opus 4.8 via claude-agent-sdk.", "totalTime": "PT15M", "estimatedCost": { "@type": "MonetaryAmount", "currency": "USD", "value": "20" }, "tool": [ { "@type": "HowToTool", "name": "Claude Overlay (MIT)" }, { "@type": "HowToTool", "name": "Claude Code CLI" }, { "@type": "HowToTool", "name": "Python 3.10+" }, { "@type": "HowToTool", "name": "claude-agent-sdk, Pillow, keyboard" } ], "step": [ { "@type": "HowToStep", "position": 1, "name": "Clone and run setup.cmd", "text": "git clone https://github.com/shengyanlin/claude-overlay.git and run setup.cmd from an Administrator PowerShell. The script installs Python packages and Claude Code CLI, then opens browser for OAuth authentication.", "url": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026#step-1", "timeRequired": "PT5M" }, { "@type": "HowToStep", "position": 2, "name": "Launch the overlay", "text": "Run python claude_overlay.py from the Administrator terminal. The Tkinter frameless window appears at the bottom-right of the primary monitor. Always-on-top, no focus steal.", "url": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026#step-2", "timeRequired": "PT1M" }, { "@type": "HowToStep", "position": 3, "name": "Configure capture mode and make a query", "text": "Toggle Snap or Auto capture mode. Type a question referencing visible screen content. The overlay captures monitors, labels them, and sends to Claude Code via claude-agent-sdk for vision-based analysis.", "url": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026#step-3", "timeRequired": "PT2M" }, { "@type": "HowToStep", "position": 4, "name": "Review response and collapse to orb", "text": "Read Claude's response in the overlay. Continue the conversation or click minimize to collapse to the draggable orb. The orb preserves the session state.", "url": "https://dailyaiworld.com/blogs/claude-overlay-screen-aware-desktop-agent-2026#step-4", "timeRequired": "PT2M" } ] } ] } SCHEMA_DATA_END

SUPABASE PAYLOAD END