Peekaboo MCP

# Anthropic SDK Integration Plan for Peekaboo ## Overview This document outlines the plan for adding Anthropic Claude support to Peekaboo, enabling users to leverage Claude models alongside existing OpenAI models. ## SDK Options Analysis ### 1. Community Swift SDKs **AnthropicSwiftSDK** (https://github.com/fumito-ito/AnthropicSwiftSDK) - ✅ Pure Swift implementation - ✅ Async/await support - ✅ Tool/function calling via macros - ❌ External dependency - ❌ May lag behind API updates **SwiftAnthropic** (https://github.com/jamesrochabrun/SwiftAnthropic) - ✅ Comprehensive API coverage - ✅ Well-documented - ✅ Multi-platform support - ❌ External dependency - ❌ Different architectural patterns than Peekaboo ### 2. Official TypeScript SDK **anthropic-sdk-typescript** (https://github.com/anthropics/anthropic-sdk-typescript) - ✅ Official support - ✅ Always up-to-date - ❌ Requires Node.js bridge - ❌ Performance overhead - ❌ Complexity of cross-language integration ### 3. Custom Implementation (Recommended) Build a native Swift implementation following Peekaboo's existing patterns. **Advantages:** - ✅ Consistent with existing OpenAI implementation - ✅ No external dependencies - ✅ Full control over features and compatibility - ✅ Optimized for Peekaboo's architecture - ✅ Direct API calls without subprocess overhead **Disadvantages:** - ❌ More initial development work - ❌ Need to maintain API compatibility ## Recommendation: Custom Swift Implementation Based on Peekaboo's philosophy of "No Backwards Compatibility" and preference for clean, modern code without dependencies, I recommend building a custom Anthropic implementation. ## Implementation Plan The plan leverages Peekaboo's existing protocol-based architecture, as demonstrated in the recent OpenAI API migration. This ensures the Anthropic integration will be a natural extension of the current system. ### Phase 1: Core Components (2-3 days) 1. **Create Anthropic-specific types** - Location: `Core/PeekabooCore/Sources/PeekabooCore/AI/Anthropic/AnthropicTypes.swift` - Request/response structures - Message format types - Tool/function definitions - Error types 2. **Create AnthropicModel.swift** - Location: `Core/PeekabooCore/Sources/PeekabooCore/AI/Models/AnthropicModel.swift` - Implement `ModelInterface` protocol (matching OpenAIModel pattern) - Basic message creation - Authentication handling - Error handling 3. **API Integration** - Endpoint: `https://api.anthropic.com/v1/messages` - Headers: `x-api-key`, `anthropic-version` - Request/response parsing ### Phase 2: Streaming Support (1-2 days) 1. **SSE Parser** - Handle server-sent events - Parse streaming chunks - Convert to Peekaboo's `StreamEvent` types 2. **Streaming Integration** - Implement `getStreamedResponse` - Handle partial content blocks - Tool streaming support ### Phase 3: Tool/Function Calling (1-2 days) 1. **Tool Conversion** - Map `ToolDefinition` to Anthropic format - Handle `tool_use` content blocks - Process `tool_result` responses 2. **Tool Streaming** - Stream tool calls - Handle partial JSON in tool arguments ### Phase 4: Integration (1 day) 1. **Model Registration** ```swift // In ModelProvider.swift registerAnthropicModels() ``` 2. **Credential Management** - Add `ANTHROPIC_API_KEY` support - Update configuration system 3. **Model Configuration** - Register Claude models: - claude-3-opus-latest - claude-3-sonnet-20240229 - claude-3-haiku-20240307 - claude-3-5-sonnet-latest - claude-sonnet-4-20250514 ### Phase 5: Testing & Polish (1-2 days) 1. **Testing** - Unit tests using Swift Testing framework - Integration tests - Error scenario testing 2. **Documentation** - Update CLAUDE.md - Add usage examples - API documentation ## Technical Details ### Message Conversion ```swift // Peekaboo → Anthropic SystemMessageItem → system parameter UserMessageItem → messages[].content AssistantMessageItem → messages[].content ToolMessageItem → tool_result content block ``` ### Authentication ```swift var request = URLRequest(url: url) request.setValue(apiKey, forHTTPHeaderField: "x-api-key") request.setValue("2023-06-01", forHTTPHeaderField: "anthropic-version") ``` ### Key Differences from OpenAI 1. **Message Format**: Anthropic uses a different message structure with content blocks 2. **System Prompt**: Separate parameter, not a message 3. **Tool Format**: Different schema for tool definitions 4. **Streaming**: Different event types and format 5. **Error Handling**: Different error response structure ## Implementation Progress ### Completed ✅ - [x] Created AnthropicTypes.swift with all request/response types - [x] Created AnthropicModel.swift implementing ModelInterface - [x] Added authentication with x-api-key header - [x] Implemented basic message creation (non-streaming) - [x] Implemented SSE parser for streaming responses - [x] Integrated streaming with getStreamedResponse - [x] Mapped Anthropic events to StreamEvent types - [x] Implemented tool definition conversion - [x] Added tool_use and tool_result content blocks - [x] Supported tool streaming events - [x] Registered all Anthropic models in ModelProvider - [x] Added ANTHROPIC_API_KEY credential handling - [x] Updated agent command to support Anthropic models - [x] Updated CLAUDE.md documentation - [x] Created initial test suite ### Remaining 🚧 - [ ] Run integration tests with actual API - [ ] Verify tool calling with real examples - [ ] Performance testing ## Success Criteria - [x] Basic message generation working - [x] Streaming responses functional - [x] Tool calling implemented - [x] All Claude models registered - [x] Tests passing - [ ] Performance on par with OpenAI implementation ## Timeline Estimated total: 7-10 days for complete implementation ## Next Steps 1. Review and approve this plan 2. Create branch for Anthropic integration 3. Begin Phase 1 implementation