DollhouseMCP

# Session Notes - Claude Code Plugin Ambient Intelligence Architecture **Date**: October 9, 2025 **Time**: Evening (6:30 PM - 10:30 PM) **Focus**: Claude Code plugin analysis and ambient intelligence architecture design **Outcome**: ✅ Discovered hooks can create AI copilot, documented complete architecture --- ## Executive Summary Transformational session that pivoted from simple slash commands to designing DollhouseMCP as an **ambient intelligence layer**. Discovered that Claude Code hooks can monitor all messages, query the Capability Index locally (0 tokens), and intelligently inject context. This enables DollhouseMCP to become a proactive AI copilot that knows everything you've built and can create anything you need. **Key Breakthrough**: The 46,000-token Capability Index and all memories exist as a local knowledge graph that hooks can query without consuming ANY tokens. A universal hook becomes the bridge between vast local knowledge and Claude's limited context window. --- ## Major Discoveries ### 1. Claude Code Plugin Announcement (October 9, 2025) Anthropic announced plugin support for Claude Code today. Initial analysis revealed: - Plugins support slash commands, agents, hooks, and MCP servers - Distribution via GitHub repositories and marketplaces - Hooks can inject `additionalContext` into Claude ### 2. The Capability Index Reality **Critical Insight**: The Capability Index is a client-side search engine, not context content - 46,000 tokens of YAML that NEVER loads into Claude - Acts as local intelligence for decision-making - Hooks query it locally (0 token cost) - Enables infinite scaling without token penalties ### 3. User's Vision Crystallizes Mick articulated the real opportunity: - One universal hook monitoring everything - Query Capability Index for relevant capabilities - Surface memories dynamically - Offer element creation when gaps detected - Transform every interaction into an assistance opportunity ### 4. Semantic Understanding Gap Identified that Anthropic isn't discussing using lightweight LLMs (like Haiku) in hooks for semantic understanding. This creates opportunity for hybrid approach: - Programmatic pattern matching (5ms, free) - Haiku/local LLM fallback (200ms, $0.000125) - Combines speed with intelligence --- ## Technical Accomplishments ### 1. Hook Research Completed Discovered hook architecture: - Configuration in `~/.claude/settings.json` - UserPromptSubmit event perfect for monitoring - Can return JSON with additionalContext - 60-second timeout, local execution - Full file system access ### 2. Ambient Intelligence Architecture Documented Created comprehensive `AMBIENT_INTELLIGENCE_ARCHITECTURE.md`: - Universal hook design pattern - Complete Python implementation - Capability Index integration - Memory search functionality - Performance metrics (90%+ token savings) ### 3. MVP Plan Developed Weekend sprint plan: - Friday: Core hook (4 hours) - Saturday: Testing and refinement - Sunday: Polish and demo - Tuesday: Launch as first ambient AI for Claude Code --- ## Key Insights ### 1. DollhouseMCP Core Value Clarified "DollhouseMCP allows anyone to make anything they need simply by asking" - Creation is the CORE feature, not an add-on - Audio summarizer created in 2 lines, used daily - Every element immediately discoverable ### 2. Competitive Moat Identified The Capability Index + hook architecture creates a 6+ month moat: - 20,559 lines of structured intelligence - Verb-to-capability mappings - Semantic relationships - Usage tracking and optimization - No one else has this infrastructure ### 3. Patent Connection Discovered Mick's 4 patents from 15 years ago on NLP for unstructured data (tweets/Facebook) directly apply: - Intent classification - Semantic clustering - Tool/action mapping These patterns are exactly what's needed for AI orchestration. --- ## Architecture Design ### Universal Hook Flow ``` User Message → Hook (local) → Query Index (0 tokens) → Inject Context (500 tokens) ``` ### Key Components 1. **Intent Extraction**: Parse verbs and purposes from messages 2. **Index Query**: Search 46K tokens locally 3. **Memory Search**: Grep through YAML memories 4. **Smart Injection**: Only relevant context to Claude ### Dual Approach Strategy 1. **Programmatic** (5ms, free) - Regex patterns - Verb matching - Known patterns 2. **Semantic** (200ms, $0.000125) - Haiku or local LLM - True understanding - Complex intent --- ## Implementation Plan ### Phase 1: MVP Weekend Sprint - Core hook implementation - Basic pattern matching - Capability Index integration - Simple commands ### Phase 2: Semantic Enhancement - Add Haiku for complex intent - Local LLM option (Llama 3.2 1B) - A/B testing between approaches ### Phase 3: Full Ambient Intelligence - Memory integration - Creation detection - Learning from usage - Predictive loading --- ## Strategic Implications ### Market Positioning - First MCP server with ambient intelligence - Transforms from tool to AI copilot - Infinite scaling without token penalties - Self-improving through creation ### User Experience Revolution - Zero cognitive load - Proactive assistance - Automatic discovery - Natural creation flow ### Technical Advantages - Local knowledge graph (0 tokens) - Hybrid intelligence (programmatic + LLM) - Patent-backed NLP approach - 6+ month competitive moat --- ## Critical Decisions Made 1. **Focus on hooks over slash commands** - More transformational 2. **Dual approach** - Programmatic baseline + semantic enhancement 3. **Weekend MVP** - Achievable with existing infrastructure 4. **Ambient intelligence positioning** - Not just a plugin, an AI copilot --- ## Next Session Priorities ### Immediate (Friday Evening) 1. Create basic UserPromptSubmit hook 2. Test Capability Index querying 3. Implement pattern matching ### Weekend Sprint 1. Build complete MVP 2. Add semantic layer 3. Create demo 4. Prepare launch materials ### Tuesday Launch 1. Announce ambient intelligence for Claude Code 2. Position as first intelligent MCP integration 3. Demonstrate creation and discovery --- ## Files Created/Modified ### Created 1. `CLAUDE_CODE_PLUGIN_IMPLEMENTATION.md` - Original plugin plan 2. `AMBIENT_INTELLIGENCE_ARCHITECTURE.md` - Revolutionary architecture 3. Issue #1307 - Plugin tracking issue 4. PR #1308 - Plugin documentation 5. `feature/claude-code-plugin-implementation-plan` branch 6. `feature/ambient-intelligence-hooks` branch ### Key Commits - `e3b0afa4` - Plugin implementation plan - `799e6441` - Ambient intelligence architecture --- ## Metrics - Session Duration: 4 hours - Documents Created: 3 major specifications - Lines of Documentation: ~1,200 - GitHub Issues: 1 created - PRs: 1 submitted - Key Insights: 5 breakthrough moments - Token Savings Discovered: 90%+ --- ## Learning & Revelations ### Technical Learnings 1. Hooks run locally with full file system access 2. Can inject context via JSON return 3. UserPromptSubmit perfect for monitoring 4. Settings.json configuration (not in plugin) ### Strategic Learnings 1. Ambient intelligence > slash commands 2. Local knowledge graph enables infinite scale 3. Creation as core differentiator 4. Semantic understanding gap in market ### Personal Revelation Mick's patent work from 15 years ago on NLP for social media directly applies to modern AI orchestration. The patterns for tweet classification are the patterns for tool selection. --- ## Quote of the Session "The Capability Index is effectively a dynamic targeted tool that points at everything that DollhouseMCP can do... we need a hook that looks at almost every single statement that somebody uses, runs in the background and then bubbles up anything appropriate." - Mick --- ## Conclusion This session transformed our understanding of what DollhouseMCP can become. Moving from simple plugin commands to an ambient intelligence layer that monitors, suggests, and creates. The weekend MVP is not just achievable but could revolutionize how users interact with AI tools. The convergence of Mick's patent experience, the Capability Index infrastructure, and Claude Code's hook system creates a perfect storm for innovation. By Tuesday, DollhouseMCP will be the first ambient AI copilot for Claude Code. **Status**: Ready for weekend sprint. Architecture documented. Path clear. --- *Session completed successfully. Ambient intelligence architecture ready for implementation.*