Peekaboo
a macOS-only MCP server that enables AI agents to capture screenshots of applications, or the entire system.
Peekaboo 🫣 - Mac automation that sees the screen and does the clicks.

Peekaboo brings high-fidelity screen capture, AI analysis, and complete GUI automation to macOS. Version 3 adds native agent flows and multi-screen automation across the CLI and MCP server.
Note: v3 is currently in beta (3.0.0-beta4) and has a few known issues; see the changelog for details.
What you get
- Pixel-accurate captures (windows, screens, menu bar) with optional Retina 2x scaling.
- Natural-language agent that chains Peekaboo tools (see, click, type, scroll, hotkey, menu, window, app, dock, space).
- Menu and menubar discovery with structured JSON; no clicks required.
- Multi-provider AI: GPT-5.1 family, Claude 4.x, Grok 4-fast (vision), Gemini 2.5, and local Ollama models.
- MCP server for Claude Desktop and Cursor plus a native CLI; the same tools in both.
- Configurable, testable workflows with reproducible sessions and strict typing.
- Requires macOS Screen Recording + Accessibility permissions (see docs/permissions.md).
Install
- macOS app + CLI (Homebrew):
brew install steipete/tap/peekaboo - MCP server (Node 22+, no global install needed):
npx -y @steipete/peekaboo
Quick start
# Capture full screen at Retina scale and save to Desktop
peekaboo image --mode screen --retina --path ~/Desktop/screen.png
# Click a button by label (captures, resolves, and clicks in one go)
peekaboo see --app Safari --json | jq -r '.data.snapshot_id' | read SNAPSHOT
peekaboo click --on "Reload this page" --snapshot "$SNAPSHOT"
# Run a natural-language automation
peekaboo "Open Notes and create a TODO list with three items"
# Run as an MCP server (Claude/Cursor)
npx -y @steipete/peekaboo
# Minimal Claude Desktop config snippet (Developer → Edit Config):
# {
# "mcpServers": {
# "peekaboo": {
# "command": "npx",
# "args": ["-y", "@steipete/peekaboo"],
# "env": {
# "PEEKABOO_AI_PROVIDERS": "openai/gpt-5.1,anthropic/claude-opus-4"
# }
# }
# }
# }
| Command | Key flags / subcommands | What it does |
|---|---|---|
| see | --app, --mode screen/window, --retina, --json | Capture and annotate UI, return snapshot + element IDs |
| click | --on <id/query>, --snapshot, --wait, coords | Click by element ID, label, or coordinates |
| type | --text, --clear, --delay-ms | Enter text with pacing options |
| press | key names, --repeat | Special keys and sequences |
| hotkey | combos like cmd,shift,t | Modifier combos (cmd/ctrl/alt/shift) |
| scroll | --on <id>, --direction up/down, --ticks | Scroll views or elements |
| swipe | --from/--to, --duration, --steps | Smooth gesture-style drags |
| drag | --from/--to, modifiers, Dock/Trash targets | Drag-and-drop between elements/coords |
| move | --to <id/coords>, --screen-index | Position the cursor without clicking |
| window | list, move, resize, focus, set-bounds | Move/resize/focus windows and Spaces |
| app | launch, quit, relaunch, switch, list | Launch, quit, relaunch, switch apps |
| space | list, switch, move-window | List or switch macOS Spaces |
| menu | list, list-all, click, click-extra | List/click app menus and extras |
| menubar | list, click | Target status-bar items by name/index |
| dock | launch, right-click, hide, show, list | Interact with Dock items |
| dialog | list, click, input, file, dismiss | Drive system dialogs (open/save/etc.) |
| image | --mode screen/window/menu, --retina, --analyze | Screenshot screen/window/menu bar (+analyze) |
| list | apps, windows, screens, menubar, permissions | Enumerate apps, windows, screens, permissions |
| tools | --verbose, --json, --no-sort | Inspect native Peekaboo tools |
| config | init, show, add, login, models | Manage credentials/providers/settings |
| permissions | status, grant | Check/grant required macOS permissions |
| run | .peekaboo.json, --output, --no-fail-fast | Execute .peekaboo.json automation scripts |
| sleep | --duration (ms) | Millisecond delays between steps |
| clean | --all-snapshots, --older-than, --snapshot | Prune snapshots and caches |
| agent | --model, --dry-run, --resume, --max-steps, audio | Natural-language multi-step automation |
| mcp | serve (default) | Run Peekaboo as an MCP server |
Models and providers
- OpenAI: GPT-5.1 (default) and GPT-4.1/4o vision
- Anthropic: Claude 4.x
- xAI: Grok 4-fast reasoning + vision
- Google: Gemini 2.5 (pro/flash)
- Local: Ollama (llama3.3, llava, etc.)
Set providers via PEEKABOO_AI_PROVIDERS or peekaboo config add.
Learn more
- Command reference: docs/commands/
- Architecture: docs/ARCHITECTURE.md
- Building from source: docs/building.md
- Testing guide: docs/testing/tools.md
- MCP setup: docs/commands/mcp.md
- Permissions: docs/permissions.md
- Ollama/local models: docs/ollama.md
- Agent chat loop: docs/agent-chat.md
- Service API reference: docs/service-api-reference.md
Development basics
- Requirements: macOS 15+, Xcode 16+/Swift 6.2. Node 22+ only if you run the pnpm docs/build helper scripts (core CLI/app/MCP are Swift-only).
- Install deps:
pnpm installthenpnpm run build:cliorpnpm run test:safe. - Lint/format:
pnpm run lint && pnpm run format.
License
MIT
Related Servers
Liquidium MCP Server
Tools for PostHog analytics, annotations, and project management.
GenieACS MCP
MCP server that exposes GenieACS TR-069 ACS instances to LLMs for device management, firmware downloads, and parameter reads
Enterpret
Enterpret's Wisdom MCP Server brings customer intelligence directly into your favorite AI tools.
Pluga
Connect your AI Agents to automation workflows as if by magic
Umbraco MCP
Interact with the Umbraco CMS Management API for administrative tasks.
Feishu/Lark OpenAPI
Connect AI agents to Feishu/Lark APIs for automating tasks like document processing, conversation management, and calendar scheduling.
Ortto MCP Server
Ortto’s new MCP Server lets you connect your favorite AI assistant to Ortto. Once connected, your assistant can securely access customer data, audiences, and campaign insights from Ortto’s CDP and marketing automation platform.
TaskMaster Todoist
An MCP server for integrating Todoist with Cursor AI. Requires a Todoist API token.
Chhart MCP
Chhart MCP is a tool that enables AI assistants to generate instant, shareable flowcharts and Sankey diagrams directly in chat,
laundry-timer-mcp
A laundry planning assistant that uses preferences and real-time weather forecasts.