ai-garden-railway-mcp

DayDreamerAI/ai-garden-railway-mcp

3.1

If you are the rightful owner of ai-garden-railway-mcp and would like to certify it and/or have it hosted online, please leave a comment on the right or send an email to henry@mcphub.com.

The AI Garden Enhanced Railway MCP Server v2.3.0 is a security-enhanced server designed for deployment on Railway, featuring comprehensive auditing and health monitoring.

Daydreamer MCP Server - Cloud Run + OAuth 2.1

Multi-Platform Memory Sovereignty: Desktop + Web + Mobile

Version: 6.7.2+oauth2.1+sse-fixes+oauth-debug (October 28, 2025) Stdio Parity: 100% (17/17 tools) Status: ✅ PRODUCTION - SSE ISSUES #10-11 RESOLVED, OAuth #12 DIAGNOSED

🚀 Live Deployment (Cloud Run - Primary)

  • Service URL: https://daydreamer-mcp-connector-480492152047.us-central1.run.app
  • SSE Endpoint: https://daydreamer-mcp-connector-480492152047.us-central1.run.app/sse
  • Platform: Google Cloud Run (serverless, auto-scaling)
  • Revision: 00033-5nl (serving 100% traffic) - Enhanced OAuth logging active
  • Protocol: SSE Transport + OAuth 2.1
  • Database: Neo4j AuraDB InstanceDaydreamer_01
  • Configuration: 8GB memory, 3600s timeout, 50 max connections, us-central1
  • Cost: $10-20/month (vs Railway $20+/month)
  • Recent Fixes:
    • Issue #10: SSE timeout aligned with OAuth (3600s)
    • Issue #11: Connection limit increased to 50
    • Issue #12: Enhanced OAuth logging for reconnection debugging (client-side bug documented)

🔐 OAuth 2.1 Authentication

Dynamic Client Registration - No manual credentials needed!

Claude Web/Mobile automatically:

  1. Discovers OAuth endpoints (/.well-known/oauth-authorization-server + /.well-known/oauth-protected-resource)
  2. Registers as OAuth client via /register endpoint
  3. Obtains authorization code with PKCE S256
  4. Exchanges code for JWT access token (1hr expiry)
  5. Connects to SSE endpoint with Bearer token

RFC Compliance:

  • ✅ RFC 7591: Dynamic Client Registration
  • ✅ RFC 7636: PKCE with S256
  • ✅ RFC 8414 Section 3: Authorization Server Metadata
  • ✅ RFC 8414 Section 5: Protected Resource Metadata
  • ✅ MCP Authorization Specification 2025-03-26

📱 Claude Web/Mobile Custom Connector Setup

Quick Setup

  1. Open Claude Web → Settings → Custom Connectors → Add Custom Connector
  2. Enter Details:
    • Name: Daydreamer Memory
    • URL: https://daydreamer-mcp-connector-6j5i7oc4na-uc.a.run.app/sse
    • ⚠️ IMPORTANT: Must include /sse endpoint path
  3. Click "Connect"
  4. OAuth Flow: Automatic (no credentials needed)
    • Client registration
    • Authorization
    • Token exchange
    • Connection established

Troubleshooting

"Disconnected" status after adding connector?

  • ✅ Verify URL includes /sse path
  • ❌ Wrong: https://daydreamer-mcp-connector-6j5i7oc4na-uc.a.run.app (base URL alone)
  • ✅ Correct: https://daydreamer-mcp-connector-6j5i7oc4na-uc.a.run.app/sse

OAuth flow details (for debugging):

  • View logs: gcloud run services logs read daydreamer-mcp-connector --region us-central1
  • Expected: Client registration → Authorization → Token → SSE connection
  • Client IDs: Auto-generated like mcp_<random>

🖥️ Claude Desktop Setup (Stdio - Alternative)

For Claude Desktop, use the stdio MCP server (no OAuth needed):

{
  "mcpServers": {
    "daydreamer-memory": {
      "command": "python",
      "args": ["/path/to/daydreamer-mcp/server/main.py"]
    }
  }
}

🔄 Railway Deployment (Legacy - Backup)

URL: https://ai-garden-railway-mcp-production.up.railway.app Status: Active (will be decommissioned after 1 week Cloud Run validation) Auth: Legacy bearer token (backward compatibility)

Production Validation

Cloud Run Deployment: Commit c5cfe4d (Oct 26, 2025)

Validated Features:

  • ✅ OAuth 2.1 complete (5 endpoints operational)
  • ✅ Dynamic Client Registration (DCR) working
  • ✅ PKCE S256 verification passing
  • ✅ JWT access tokens (HS256, 1hr expiry)
  • ✅ Claude Web custom connector connected
  • ✅ PBC Desktop V2.0.12 entity loading (Julian + Claude with observations)
  • ✅ Observation source property (provenance tracking)
  • ✅ search_observations Cypher fix (ORDER BY aggregation)
  • ✅ GraphRAG global search (241 Leiden communities)
  • ✅ GraphRAG vector index (community_summary_vector_idx ONLINE)

Platforms Confirmed: Claude Desktop (stdio), Claude Web (OAuth + SSE), Mobile (OAuth + SSE)

🛠️ Complete Tool Suite (17 Tools - 100% Stdio Parity)

Entity & Observation Management (5 tools)

  1. create_entities - Create entities with V6 observations, JinaV3 embeddings, MVCM
  2. add_observations - Add observations with direct Cypher, semantic themes, MVCM entity mentions
  3. search_nodes - Semantic search via JinaV3 256D or exact name/alias lookup
  4. create_relations - Create entity relationships with canonical types
  5. search_observations - Multi-dimensional filtering (theme, entity, date, confidence)

Conversation & Temporal Tools (4 tools)

  1. search_conversations - Search preserved conversation sessions by topic/date/activity
  2. trace_entity_origin - Find which conversations created/discussed an entity
  3. get_temporal_context - Get conversations within time window around specific dates
  4. get_breakthrough_sessions - Retrieve high-importance conversation sessions

GraphRAG Advanced Search (2 tools)

  1. graphrag_global_search - Community-level synthesis across 241 Leiden clusters
  2. graphrag_local_search - Entity neighborhood exploration with multi-hop traversal

System Management (3 tools)

  1. memory_stats - System health, embedding coverage, V6 compliance metrics
  2. raw_cypher_query - Direct Neo4j Cypher access for operational queries
  3. generate_embeddings_batch - JinaV3 256D batch embedding generation (Railway cloud-native)

Advanced Memory Tools (3 tools - Stubs)

  1. conversational_memory_search - Natural language memory search (fallback to enhanced_search)
  2. virtual_context_search - 70% token reduction search (fallback to regular search)
  3. lightweight_embodiment - <4K token startup protocol (recommend full PBC via Claude Code)

📊 Production Data

  • Entities: 27,487
  • Relationships: 110,449
  • V6 Observations: 14,414
  • Conversation Sessions: 537

🔧 Environment Variables

Required environment variables (set in Railway):

NEO4J_URI=neo4j+s://8c3b5488.databases.neo4j.io
NEO4J_PASSWORD=<your-auradb-password>
JINA_API_KEY=<your-jina-api-key>

Optional Performance Variables

EMBEDDING_TIMEOUT=40        # Default: 40 seconds (configured for Railway CPU)
ENABLE_AUTO_UNLOAD=false    # Default: false (keep model resident)

EMBEDDING_TIMEOUT: Maximum time to wait for embedding generation before fallback.

  • Default: 40 seconds (allows ~24s CPU model load + safety buffer)
  • Configured: Optimized for Railway CPU lazy loading

ENABLE_AUTO_UNLOAD: Whether to unload JinaV3 model after idle timeout.

  • Default: false (model stays resident once loaded)
  • Benefit: Instant subsequent queries, no reload needed
  • Memory: ~3.7GB persistent (acceptable for Railway environment)

⚡ Performance Characteristics

JinaV3 Lazy Loading (CPU Environment)

First Embedding Request (~24 seconds):

  • Model downloads from HuggingFace (if not cached)
  • 3.2GB model loads into CPU memory
  • Int8 quantization applied
  • 256D Matryoshka truncation configured

Subsequent Requests (instant):

  • Model stays resident in memory
  • No reload required until Railway container restarts
  • Fast inference on all embedding operations

Memory Profile:

  • Startup: ~500MB (no model loaded)
  • After first embedding: ~3.7GB (model + base server)
  • Peak: ~4.2GB (during embedding generation)

Why Lazy Loading?:

  • Prevents 6.28GB startup memory spike (exceeded Railway's 4.5GB threshold)
  • Enables mobile/web connections without memory circuit breaker
  • Optimal for read-only sessions (never loads model if no embeddings needed)
  • Implemented in v6.3.3 (October 19, 2025)

📦 Deployment

Railway auto-deploys from main branch.

Manual Deployment

# Install Railway CLI
brew install railway

# Login
railway login

# Deploy
railway up

🧪 Testing

# Test production endpoint
curl https://ai-garden-railway-mcp-production.up.railway.app/health

# Test search (triggers JinaV3 lazy load on first request)
# Expect: ~24s first time, instant subsequent requests

📚 Documentation

Quick Links

  • - Complete version history (v6.3.6: 95% V6 compliance)
  • - V6 compliance audits and responses
    • Latest: 19/20 (95%) - October 19, 2025
    • All fixes documented and validated
  • - Technical documentation and deployment guides

System Status

  • V6 Compliance: 19/20 (95%) ✅ PRODUCTION READY
  • Architecture: Custom Connector + SSE transport
  • Memory System: V6 observation nodes with temporal binding
  • GraphRAG: Phase 3 complete (241 communities, global/local search operational)

🆕 Recent Updates

v6.7.1 (October 22, 2025) - 🔧 STDIO SYNC - Observation Source Property:

  • ✅ Source Property Added: Observations now include source property for provenance tracking
  • ✅ Stdio Parity Maintained: Matches stdio v6.7.0 post-Oct 21 fixes
  • ✅ Backward Compatible: Defaults to "manual-reflection" for existing workflows
  • ✅ Schema Compliant: Full V6 observation property compliance (id, content, created_at, semantic_theme, conversation_id, source)

v6.7.0 (October 20, 2025) - 🎯 COMPLETE STDIO PARITY - 47% → 100% Tool Coverage:

  • ✅ V6 Bridge Removed: Deprecated V6MCPBridge replaced with direct Cypher implementation (stdio v6.6.0)
  • ✅ Direct Cypher: Stdio-identical observation creation with MVCM entity mentions
  • ✅ MVCM Integration: Automatic entity mention detection via concept extraction (stdio v6.7.0)
  • ✅ 9 New Tools: create_relations, search_observations, 4 conversation tools, 3 advanced stubs
  • ✅ Tool Parity: 8 tools → 17 tools (100% stdio v6.7.0 feature parity)
  • ✅ ONE CANONICAL LOGIC: Railway = stdio + SSE transport (zero logic variations)
  • ✅ Schema Compliance: Updated paths to /llm/memory/standards/MEMORY_V6_COMPLIANCE_STANDARDS.md
  • ✅ SemanticThemeClassifier: 9 themes + MVCM concept extraction integrated directly
  • ✅ Localhost Protection: Triple-layer security preventing local connections (stdio v5.1.0)

v6.3.6 (October 19, 2025) - ✅ V6 AUDIT FIXES COMPLETE - 70% → 95% Compliance:

  • ✅ V5 Chunk Migration: 3,428 nodes migrated to V6 schema (100% compliance)
  • ✅ Observation Reclassification: 9,343 observations reclassified (97.5% → 49.2% "general")
  • ✅ Community Membership: Audit bug resolved - actual coverage 99.7% (exceeds 80% target)
  • ✅ V6 Compliance: Improved from 70% (14/20) to 95% (19/20 requirements)
  • ✅ Documentation: AUDIT_FIX_COMPLETE.md with corrected audit queries

v6.3.5 (October 19, 2025) - 🔧 SSE Connection Management + V5 Migration:

  • ✅ SSE Connections: Increased limit from 5→10 for multi-platform usage
  • ✅ Auto-Cleanup: 5-minute timeout + background stale session cleanup
  • ✅ V5 Migration Tool: Script to migrate 3,427 legacy Chunk nodes to V6 schema
  • ✅ Better Monitoring: Session tracking with timestamps and active connection logging

v6.3.4 (October 19, 2025) - 🔥 CRITICAL FIXES - MCP Protocol + GraphRAG Global Search:

  • ✅ MCP Protocol Compliance: Fixed "Unknown method: prompts/list" error (added 4 missing handlers)
  • ✅ GraphRAG Global Search: Fixed embedder isolation causing memory spikes and failures
  • ✅ True Lazy Loading: Removed warmup call that defeated v6.3.3 lazy loading (startup memory now ~500MB)
  • ✅ Memory Stability: Global search uses singleton embedder (no duplicate 3.2GB model loading)

v6.3.3 (October 19, 2025) - 🔥 CRITICAL FIX - True Lazy Loading (incomplete):

  • ✅ JinaV3 Model Lazy Loading: Fixed model loading at startup (was consuming 3.2GB immediately)
  • ✅ Startup Memory: Reduced from 6.28GB to ~3.0GB (below 4.5GB circuit breaker threshold)
  • ✅ Mobile Connections Restored: Model only loads when actually needed (may never load if read-only)
  • ✅ Memory Monitoring Disabled: Eliminated misleading system memory logs

v6.3.2 (October 18, 2025) - V6 Compliance & Resource Optimization:

  • ✅ Semantic Theme Classification: Fixed 97.5% "general" theme bug with lightweight classifier
  • ✅ V5 Timestamp Removal: 100% V6 compliance with ISO 8601 created_at properties
  • ✅ Resource Optimization: CPU-only mode for Railway environment
  • ✅ Memory Efficiency: Lazy initialization, pattern-based classification

Impact: Production stability restored, proper theme distribution, V6 schema compliance

Deployment Version: Tier 1 (8 tools) | v6.3.5 Last Updated: October 19, 2025 Source: Private daydreamer-mcp repository