The Peekaboo MCP server provides macOS screen capture and AI-powered image analysis for integration with AI assistants like Claude Desktop and Cursor IDE.
Capture Screenshots: Take screenshots of entire screens, specific app windows, or all windows of an app, with control over output format (file or Base64 data).
AI Image Analysis: Analyze captured or existing images using AI models like GPT-4 Vision, Claude, or local models via Ollama.
System Information: List running applications, their open windows (including details like position, size, and IDs), and server status.
Integration: Seamlessly integrate with AI assistants for visual context and debugging workflows.
Privacy Options: Support for local AI analysis through Ollama for privacy-focused use cases.
Advanced Targeting: Fuzzy matching and precise targeting for apps, windows, and screens.
Enables comprehensive screen capture capabilities on macOS, including capturing entire screens, specific application windows, or all windows of an app with various formatting options.
Enables local AI image analysis of screenshots through Ollama, supporting models like LLaVA and Qwen2-VL for vision tasks without sending data to the cloud.
Provides integration with OpenAI's vision models (like GPT-4o) for analyzing captured screenshots through the OpenAI API.
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Peekaboo MCP take a screenshot of my current screen and describe what you see"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Peekaboo is a macOS CLI and menu-bar app for screen capture, accessibility inspection, and native UI automation. Use it directly, let its agent plan multi-step work, or expose the same toolset to MCP clients.
The released CLI and app require macOS 15 or later.
brew install steipete/tap/peekabooThe npm package requires Node.js 22 or later and includes the CLI plus its MCP launcher.
npx -y @steipete/peekaboo --versionSee MCP setup to connect it to Codex, Claude Code, Cursor, or another MCP client.
Download the signed DMG from the latest GitHub release. The menu-bar app provides permission onboarding, visual feedback, and agent sessions; install the CLI separately when you also need peekaboo on PATH.
For source builds and alternative install details, see the installation guide.
Related MCP server: Screen View MCP
Check the permissions available to Peekaboo, then take a screenshot:
peekaboo permissions status
peekaboo see --no-elements --mode screen --path /tmp/peekaboo-screen.pngScreen capture requires Screen Recording permission. Accessibility permission enables UI inspection and control; the permissions guide covers setup and the additional permission used for synthetic input.
Inspect a running app to get a structured UI map with opaque element IDs:
peekaboo see --app Finder --jsonThat is the core loop: observe the current screen, choose an element from the result, and act on it.
Peekaboo 4.2.3 keeps credentials out of process lists with secure prompts, stdin, and owner-only files while hardening Gemini, OAuth, clipboard, and editor workflows. Window inventory now explains whether combined Accessibility capture, screenshot-only recovery, or refreshed evidence is available. Background automation supports verified non-modal SwiftUI windows, isolates exact targets, and exposes only policy-safe Agent and MCP tools. Faster caller-local startup, Bridge-bound capture, and clearer browser, help, locked-session, and window-close recovery make everyday automation more predictable.
List Safari's windows, copy the intended window_id (12345 below), then keep the entire interaction pinned to
that exact window:
peekaboo window list --app Safari --json
peekaboo click "Address and search bar" --app Safari --window-id 12345
peekaboo type "github.com/openclaw/Peekaboo" --app Safari --window-id 12345
peekaboo press Return --app Safari --window-id 12345Targeted semantic and typed CLI input uses background delivery when Peekaboo can resolve the process, so the app does
not have to become frontmost. Raw CLI press chords can also stay background with an exact window selector. The CLI
also accepts a fresh exact non-dialog snapshot; that snapshot is required by background-only Agent/MCP policy.
App/PID-only and targetless chords require explicit foreground consent, as do window-selector-only Agent/MCP chords.
Prefer a semantic action such as menu click when one exists. See the
automation guide for element IDs, coordinates, snapshots, waits, and input behavior.
The agent combines the same observation and action tools into a natural-language run:
peekaboo agent "Open Safari, go to github.com, and search for Peekaboo" --allow-foregroundAgent runs need a configured model provider. See agent setup for providers and sessions, or MCP setup to expose Peekaboo's tools to another client.
Goal | Commands | Guide |
Observe the desktop |
| |
Interact with UI |
| |
Control macOS |
| |
Run workflows |
| |
Integrate with clients |
|
Run peekaboo help <command> for live CLI help. The complete command index links to flags, examples, and troubleshooting for every command.
Peekaboo stores provider credentials and settings under ~/.peekaboo. Use peekaboo config to inspect or change them, and consult the configuration guide for profiles, environment variables, and custom providers. The provider reference covers hosted, compatible, and local model backends.
Shell completions for zsh, bash, and fish come from peekaboo completions; see the completion guide for persistent setup.
PeekabooWin — Windows-first rewrite of the Peekaboo automation loop (JavaScript + PowerShell) by @FelixKruger
PeekabooX — Linux-first rewrite of the Peekaboo automation loop (Rust + Python) by @nordbyte
Source builds require macOS 15 or later, Swift 6.2 or later, Node.js 22 or later, and the repository's submodules.
pnpm install --frozen-lockfile
pnpm run build:cli
pnpm run lint:docs
pnpm run test:safeMore build, signing, and test details live in docs/building.md.
MIT. See LICENSE.
Analyzes a pre-existing image file from the local filesystem using a configured AI model.
This tool is useful when an image already exists (e.g., previously captured, downloaded, or generated) and you need to understand its content, extract text, or answer specific questions about it.
Capabilities:
Image Understanding: Provide any question about the image (e.g., "What objects are in this picture?", "Describe the scene.", "Is there a red car?").
Text Extraction (OCR): Ask the AI to extract text from the image (e.g., "What text is visible in this screenshot?").
Flexible AI Configuration: Can use server-default AI providers/models or specify a particular one per call via 'provider_config'.
Example: If you have an image '/tmp/chart.png' showing a bar chart, you could ask: { "image_path": "/tmp/chart.png", "question": "Which category has the highest value in this bar chart?" } The AI will analyze the image and attempt to answer your question based on its visual content. Peekaboo MCP 1.1.0 using openai/gpt-4-vision, anthropic/claude-3-opus, google/gemini-pro-vision
| Name | Required | Description | Default |
|---|---|---|---|
| image_path | No | Required. Absolute path to image file (.png, .jpg, .webp) to be analyzed. | |
| provider_config | No | Optional. Explicit provider/model. Validated against server's PEEKABOO_AI_PROVIDERS. | |
| question | Yes | Required. Question for the AI about the image. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries full burden and does well by explaining capabilities (image understanding, OCR, flexible AI configuration), supported file types, and example usage. It mentions server configuration dependencies but could be more explicit about potential limitations like image size constraints or rate limits.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is appropriately sized and well-structured with clear sections: purpose statement, usage context, capabilities list, and example. The capabilities section could be more concise, but overall it's front-loaded with essential information and every sentence adds value.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a tool with 3 parameters, 100% schema coverage, but no annotations or output schema, the description does well by explaining capabilities, usage context, and providing concrete examples. It could be more complete by mentioning potential error cases or response format expectations, but covers most essential aspects given the complexity.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so the baseline is 3. The description adds some value by explaining the purpose of parameters in context (e.g., 'question' parameter examples, 'provider_config' flexibility), but doesn't provide significant additional semantics beyond what's already well-documented in the schema.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool analyzes pre-existing image files using AI models, specifying the resource (image files) and verb (analyze). It distinguishes from sibling tools 'image' and 'list' by focusing on AI-powered analysis rather than basic image operations or listing functions.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description explicitly states when to use this tool: 'when an image already exists (e.g., previously captured, downloaded, or generated) and you need to understand its content, extract text, or answer specific questions about it.' It provides clear context for usage without needing to mention specific exclusions since the scope is well-defined.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Captures macOS screen content and optionally analyzes it. Targets can be entire screen, specific app window, or all windows of an app (via app_target). Supports foreground/background capture. Output via file path or inline Base64 data (format: "data"). If a question is provided, image is analyzed by an AI model (auto-selected from PEEKABOO_AI_PROVIDERS). Window shadows/frames excluded. Peekaboo MCP 1.1.0 using openai/gpt-4-vision, anthropic/claude-3-opus, google/gemini-pro-vision
| Name | Required | Description | Default |
|---|---|---|---|
| app_target | No | Optional. Specifies the capture target. For example: Omit or use an empty string (e.g., `''`) for all screens. Use `'screen:INDEX'` (e.g., `'screen:0'`) for a specific display. Use `'frontmost'` for all windows of the current foreground application. Use `'AppName'` (e.g., `'Safari'`) for all windows of that application. Use `'PID:PROCESS_ID'` (e.g., `'PID:663'`) to target a specific process by its PID. Use `'AppName:WINDOW_TITLE:Title'` (e.g., `'TextEdit:WINDOW_TITLE:My Notes'`) for a window of 'AppName' matching that title. Use `'AppName:WINDOW_INDEX:Index'` (e.g., `'Preview:WINDOW_INDEX:0'`) for a window of 'AppName' at that index. Ensure components are correctly colon-separated. | |
| capture_focus | No | Optional. Focus behavior. 'auto' (default): bring target to front only if not already active. 'background': capture without altering window focus. 'foreground': always bring target to front before capture. | auto |
| format | Yes | Optional. Output format. Can be `'png'`, `'jpg'`, `'jpeg'` (alias for jpg), or `'data'`. Format is case-insensitive (e.g., 'PNG', 'Png', 'png' are all valid). If `'png'` or `'jpg'`, saves the image to the specified `path`. If `'data'`, returns Base64 encoded PNG data inline in the response. If `path` is also provided when `format` is `'data'`, the image is saved (as PNG) AND Base64 data is returned. Defaults to `'data'` if `path` is not given. Invalid format values automatically fall back to 'png'. | |
| path | Yes | Optional. Base absolute path for saving the image. Relevant if `format` is `'png'`, `'jpg'`, or if `'data'` is used with the intention to also save the file. If a `question` is provided and `path` is omitted, a temporary path is used for image capture, and this temporary file is deleted after analysis. | |
| question | No | Optional. If provided, the captured image will be analyzed by an AI model. The server automatically selects an AI provider from the `PEEKABOO_AI_PROVIDERS` environment variable. The analysis result (text) is included in the response. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries full burden and does well by disclosing key behavioral traits: foreground/background capture options, output methods (file path or Base64), AI analysis when question provided, window shadow/frame exclusion, and auto-selection of AI providers. It misses some details like rate limits or specific error conditions.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is appropriately sized and front-loaded with core functionality, though it could be slightly more concise. Every sentence adds value: first states purpose, second details targets, third covers capture modes, fourth explains output, fifth describes analysis, sixth notes exclusions, seventh lists AI providers. Minor redundancy with schema exists.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a 5-parameter tool with no annotations and no output schema, the description provides good coverage of functionality, behavior, and context. It explains capture targets, focus behavior, output formats, and AI analysis integration. It could improve by detailing response structure or error handling, but it's largely complete for the tool's complexity.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so the schema already documents all 5 parameters thoroughly. The description mentions parameters like 'app_target', 'format', and 'question' but doesn't add significant meaning beyond what the schema provides. Baseline 3 is appropriate when schema does heavy lifting.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool captures macOS screen content and optionally analyzes it, specifying targets (entire screen, specific app window, all windows of an app) and output methods. It distinguishes from sibling tools 'analyze' and 'list' by focusing on capture functionality with optional analysis, not pure analysis or listing.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides clear context for when to use the tool (capturing screen content with optional AI analysis) and implies alternatives through sibling tools, but doesn't explicitly state when not to use it or directly compare to 'analyze' or 'list'. It gives practical guidance on target selection and output formats.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Lists various system items on macOS, providing situational awareness.
Capabilities:
Running Applications: Get a list of all currently running applications (names and bundle IDs).
Application Windows: For a specific application (identified by name or bundle ID), list its open windows.
Details: Optionally include window IDs, bounds (position and size), and whether a window is off-screen.
Multi-window apps: Clearly lists each window of the target app.
Server Status: Provides information about the Peekaboo MCP server itself (version, configured AI providers).
Use Cases:
Agent needs to know if 'Photoshop' is running before attempting to automate it. { "item_type": "running_applications" } // Agent checks if 'Photoshop' is in the list.
Agent wants to find a specific 'Notes' window to capture. { "item_type": "application_windows", "app": "Notes", "include_window_details": ["ids", "bounds"] } The agent can then use the window title or ID with the 'image' tool. Peekaboo MCP 1.1.0 using openai/gpt-4-vision, anthropic/claude-3-opus, google/gemini-pro-vision
| Name | Required | Description | Default |
|---|---|---|---|
| app | No | Required when `item_type` is `application_windows`. Specifies the target application by its name (e.g., "Safari", "TextEdit"), bundle ID, or process ID (e.g., "PID:663"). Fuzzy matching is used for names, so partial names may work. | |
| include_window_details | Yes | Optional, only applicable when `item_type` is `application_windows`. Specifies additional details to include for each window. Provide an array of strings. Example: `["bounds", "ids"]`. - `ids`: Include window ID. - `bounds`: Include window position and size (x, y, width, height). - `off_screen`: Indicate if the window is currently off-screen. | |
| item_type | Yes | Specifies the type of items to list. If omitted or empty, it defaults to 'application_windows' if 'app' is provided, otherwise 'running_applications'. Valid options are: - `running_applications`: Lists all currently running applications. - `application_windows`: Lists open windows for a specific application. Requires the `app` parameter. - `server_status`: Returns information about the Peekaboo MCP server. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries the full burden of behavioral disclosure and does so effectively. It explains the tool's capabilities in detail, mentions fuzzy matching for app identification, describes optional window details, and specifies default behaviors (item_type inference). However, it doesn't mention potential limitations like performance impact, permission requirements, or error conditions, preventing a perfect score.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is well-structured with clear sections (Capabilities, Use Cases) and uses bullet points effectively. While somewhat lengthy, every sentence serves a purpose: explaining capabilities, providing examples, and demonstrating workflow integration. The version information at the end could be considered extraneous but doesn't significantly detract from the overall clarity.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a tool with 3 parameters, 100% schema coverage, but no output schema or annotations, the description provides substantial context. It explains all three item_type options with practical examples, demonstrates parameter combinations, and shows how results feed into other tools. The main gap is the lack of output format description, which would be helpful given the absence of an output schema.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
With 100% schema description coverage, the baseline is 3. The description adds some value by providing concrete examples in the Use Cases section (e.g., 'Photoshop', 'Notes') and explaining the practical meaning of window details like 'bounds' for image capture. However, it doesn't significantly enhance parameter understanding beyond what's already documented in the comprehensive schema descriptions.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's purpose: 'Lists various system items on macOS, providing situational awareness.' It specifies three distinct capabilities (running applications, application windows, server status) and distinguishes itself from sibling tools like 'analyze' and 'image' by focusing on listing rather than analysis or image capture. The description goes beyond the name 'list' to explain what specific items can be enumerated.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides explicit usage guidance through the 'Capabilities' and 'Use Cases' sections. It clearly indicates when to use each item_type: checking if an app is running before automation, finding specific windows for image capture, and getting server status. The description also references sibling tools ('image') for follow-up actions, creating a clear workflow context.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Each tool has a clearly distinct purpose with no overlap: 'analyze' processes existing image files, 'image' captures screen content, and 'list' provides system information. The boundaries are well-defined, making it easy for an agent to select the right tool for each task.
The naming is mixed: 'analyze' and 'list' are simple verbs, while 'image' is a noun, breaking a consistent pattern. However, the names are still readable and descriptive of their functions, though they lack a uniform verb_noun or other predictable convention.
With 3 tools, the count is well-scoped for a macOS screen capture and analysis server. Each tool serves a distinct, essential function (analysis, capture, and system listing), and there are no extraneous tools, making the set focused and efficient.
The tool set covers core workflows for screen capture and analysis on macOS, including listing apps/windows, capturing images, and analyzing them. A minor gap is the lack of tools for managing or deleting captured images, but agents can work around this using the provided tools effectively.
Unclaimed servers have limited discoverability.
If you are the server author, to access and configure the admin panel.
Capture screenshots, detect visual regressions between page versions, and analyze with AI.
Let ChatGPT, Claude & Cursor use your Mac: email, calendar, iMessage, Teams, files. Local, free.
Desktop and mobile website screenshots plus page context for AI agents and automation workflows.
Privacy-first work tracking with summaries, reports, coaching, and AI-ready long-term memory.
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/openclaw/Peekaboo'
If you have feedback or need assistance with the MCP directory API, please join our Discord server
