Peekaboo/CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Common Commands

Building

# Build TypeScript server
npm run build

# Build Swift CLI only
npm run build:swift

# Build everything (Swift CLI + TypeScript)
npm run build:all

# Build universal Swift binary with optimizations
./scripts/build-swift-universal.sh

Testing

# Run all tests
npm test

# Run tests with coverage
npm run test:coverage

# Run tests in watch mode
npm run test:watch

# Run Swift tests
npm run test:swift

# Full integration test suite
npm run test:integration

Development

# Start TypeScript compilation in watch mode
npm run dev

# Run the server directly
npm start

# Clean build artifacts
npm run clean

# Lint Swift code
npm run lint:swift

# Format Swift code
npm run format:swift

Testing the MCP server

# Test with a simple JSON-RPC request
echo '{"jsonrpc": "2.0", "id": 1, "method": "tools/list"}' | node dist/index.js

# Run the MCP server (after building)
peekaboo-mcp

Code Architecture

Project Structure

• Node.js MCP Server (src/): TypeScript implementation of the Model Context Protocol server

  • index.ts: Main server entry point with MCP initialization
  • tools/: Tool implementations (image.ts, analyze.ts, list.ts)
  • utils/: Utilities for Swift CLI integration, AI providers, and server status
  • types/: Shared TypeScript type definitions

• Swift CLI (peekaboo-cli/): Native macOS binary for system interactions

  • Handles all screen capture, window management, and application listing
  • Outputs structured JSON when called with --json-output
  • Does NOT interact with AI providers directly

Key Design Patterns

1. Tool Handler Pattern: Each MCP tool follows a consistent pattern:

   • Validate input with Zod schema
   • Construct Swift CLI command
   • Execute Swift CLI and capture JSON output
   • Parse response and handle errors
   • Return MCP-formatted response

2. AI Provider Abstraction: The analyze tool supports multiple AI providers:

   • Configured via PEEKABOO_AI_PROVIDERS environment variable
   • Format: provider/model,provider/model (e.g., ollama/llava:latest,openai/gpt-4-vision-preview)
   • Auto-selection tries providers in order until one is available

3. Error Handling: Standardized error codes from Swift CLI:

   • PERMISSION_DENIED_SCREEN_RECORDING
   • PERMISSION_DENIED_ACCESSIBILITY
   • APP_NOT_FOUND
   • AMBIGUOUS_APP_IDENTIFIER
   • WINDOW_NOT_FOUND
   • CAPTURE_FAILED
   • FILE_IO_ERROR

4. Logging Strategy:

   • Uses Pino logger to file (default: /tmp/peekaboo-mcp.log)
   • No stdout logging to avoid interfering with MCP protocol
   • Debug logs from Swift CLI captured in JSON debug_logs array

Important Implementation Details

1. Universal Binary: The Swift CLI is built as a universal binary (arm64 + x86_64) for maximum compatibility

2. Permissions:

   • Screen Recording permission required for all capture operations
   • Accessibility permission only needed for foreground window focus

3. Image Capture: Always excludes window shadows/frames using CGWindowImageOption.boundsIgnoreFraming

4. Fuzzy App Matching: Swift CLI implements intelligent fuzzy matching for application names

5. Environment Variables:

   • PEEKABOO_AI_PROVIDERS: Configure AI backends for analysis
   • PEEKABOO_LOG_LEVEL: Control logging verbosity (trace, debug, info, warn, error, fatal)
   • PEEKABOO_DEFAULT_SAVE_PATH: Default location for captured images
   • PEEKABOO_CLI_PATH: Override bundled Swift CLI path

Common Development Tasks

• When modifying tool schemas, update both the Zod schema in TypeScript and ensure the Swift CLI output matches
• After Swift CLI changes, rebuild with npm run build:swift and test JSON output manually
• Use PEEKABOO_LOG_LEVEL=debug for detailed debugging during development
• Test permissions by running ./peekaboo list server_status --json-output