smian0/graphiti-mcp

Graphiti MCP with Ollama Integration

Self-contained Graphiti MCP server using Ollama cloud models for local, private knowledge graph operations.

Quick Start

cd /Users/smian/Services/graphiti-mcp

# Start services
./scripts/start.sh

# Stop services
./scripts/stop.sh

# Check status
./scripts/status.sh

# View logs
./scripts/logs.sh

Configuration

Models Used:

• LLM: GLM-4.6:cloud (via Ollama)
• Embeddings: nomic-embed-text (via Ollama)
• Database: FalkorDB (Redis-based graph database)

Endpoints:

• MCP Server: http://localhost:8000/mcp/
• Health Check: http://localhost:8000/health
• FalkorDB UI: http://localhost:3100

How It Works

Ollama Integration

Graphiti MCP requires OpenAI-compatible LLM clients. The challenge:

• Standard OpenAI provider uses beta responses.parse() API → Ollama doesn't support
• Solution: Custom openai_generic provider using standard chat.completions API

Automatic Patching

The scripts/patch-and-start.sh script automatically:

1. Installs Groq support (for provider patterns)
2. Adds OpenAIGenericClient import to factory
3. Registers openai_generic provider case
4. Starts MCP server with Ollama configuration

This happens on every container start - patches are ephemeral but automatically reapplied.

Configuration Files

config.yaml

llm:
  provider: "openai_generic"  # Uses standard chat.completions API
  model: "glm-4.6:cloud"
  providers:
    openai:
      api_key: "ollama-dummy-key"
      api_url: "http://host.docker.internal:11434/v1"

embedder:
  provider: "openai"
  model: "nomic-embed-text"
  providers:
    openai:
      api_key: "ollama-dummy-key"
      api_url: "http://host.docker.internal:11434/v1"

database:
  provider: "falkordb"

server:
  group_id: "main"
  transport: "http"

docker-compose.yml

• Image: zepai/knowledge-graph-mcp:latest
• Volumes: Config, data, and patch scripts
• Network: Uses host.docker.internal to reach Ollama on host

.mcp.json (Claude Code)

{
  "mcpServers": {
    "graphiti": {
      "type": "http",
      "url": "http://localhost:8000/mcp/",
      "description": "Graphiti temporal knowledge graph"
    }
  }
}

Architecture

┌─────────────────────────────────────────────────┐
│ Claude Code (MCP Client)                        │
│ /Users/smian/dotfiles/.mcp.json                 │
└────────────┬────────────────────────────────────┘
             │ HTTP MCP Protocol
             ▼
┌─────────────────────────────────────────────────┐
│ Graphiti MCP Server (Docker)                    │
│ • Port 8000: MCP endpoint (/mcp/)               │
│ • Patched with openai_generic provider          │
│ • Config: /app/config.yaml                      │
└────┬───────────────────────────┬────────────────┘
     │                            │
     │ OpenAI-compatible API      │ Graph storage
     ▼                            ▼
┌─────────────────────┐  ┌─────────────────────────┐
│ Ollama (Host)       │  │ FalkorDB (Container)    │
│ Port 11434          │  │ Port 6379               │
│ • GLM-4.6:cloud     │  │ Redis-based graph DB    │
│ • nomic-embed-text  │  │ UI: localhost:3100      │
└─────────────────────┘  └─────────────────────────┘

Troubleshooting

Check Server Status

curl http://localhost:8000/health
# Expected: {"status":"healthy","service":"graphiti-mcp"}

Verify Ollama Models

ollama list | grep -E "glm-4.6:cloud|nomic-embed-text"

Check Logs for Ollama Calls

docker-compose logs graphiti-mcp | grep "HTTP Request.*11434"

Common Issues

❌ "OpenAI provider configuration not found"

• Config file not mounted correctly
• Check: docker exec graphiti-mcp-server cat /app/config.yaml

❌ "404 page not found" from Ollama

• Using wrong provider (should be openai_generic, not openai)
• Patches not applied - check startup logs for "Applying Ollama compatibility patches"

❌ "No valid session ID provided"

• Restart MCP client (Claude Code) after server restarts

Development

Update Configuration

1. Edit config.yaml
2. Run docker-compose restart
3. Restart MCP client (Claude Code)

Switch Models

Update config.yaml and restart:

llm:
  model: "qwen2.5:14b"  # Any Ollama model
embedder:
  model: "mxbai-embed-large"  # Any Ollama embedding model

Inspect Patches

docker exec graphiti-mcp-server cat /app/mcp/src/services/factories.py | grep -A 20 'case "openai_generic"'

Data Persistence

• Graph Data: ./data/falkordb/ (persists across restarts)
• Episodes: Stored in FalkorDB
• Patches: Applied automatically on startup (not persisted in image)

Performance

• GLM-4.6:cloud: ~2-5s per entity extraction
• nomic-embed-text: ~100ms per embedding
• FalkorDB: 496x faster than Neo4j for graph queries

References

• Graphiti GitHub
• Graphiti MCP
• Ollama
• FalkorDB

Last Updated

2025-11-13

Status: ✅ Working - Ollama integration complete with automatic patching