Deploy and Host Graphiti MCP with Neo4j on Railway

Graphiti is the open-source temporal knowledge graph engine from Zep. Its MCP server turns that graph into long-term memory for AI agents: add_memory ingests text, chat messages, or JSON and extracts entities and relationships with bi-temporal validity, and search_memory_facts and search_nodes retrieve them with hybrid semantic, keyword, and graph search. Any MCP client, including Claude Code, Cursor, VS Code, and Claude Desktop, can use it as persistent memory across sessions.

About Hosting Graphiti MCP

Hosting the Graphiti MCP server means running the official zepai/knowledge-graph-mcp:1.1.0-standalone image next to a Neo4j 5.26 Community database that keeps the graph on a persistent volume. This template pins both images, selects the Neo4j backend and an IPv6 dual-stack bind through the server's own configuration overrides (no custom start command or config file mount), wires the two services together over Railway private networking, generates the Neo4j password, and uses /health and Neo4j's HTTP discovery endpoint for health checks. The MCP endpoint is served over streamable HTTP at /mcp on a public domain so remote MCP clients can connect. Graphiti needs an LLM and an embedder to extract knowledge, so an OPENAI_API_KEY is required at deploy time.

Common Use Cases

Agent memory: give a coding or chat agent a memory that survives sessions and keeps track of when facts became true or stale
User preference and requirement tracking: Graphiti's default entity types include Preference, Requirement, Procedure, Organization, and Event
Team knowledge graph: ingest meeting notes, tickets, and documents and query relationships instead of chunks
Multi-agent shared context: several agents read and write the same graph, partitioned by group_id
Graph-backed RAG: combine semantic retrieval with relationship traversal for better grounding

Dependencies for Graphiti MCP Hosting

Neo4j 5.26 Community: graph storage (included, with a /data volume)
OpenAI API key: required at boot for the LLM, embeddings, and reranking
Optional providers: Anthropic, Gemini, Groq, or Azure OpenAI for the LLM and Gemini or Voyage for embeddings, selected with LLM__PROVIDER and EMBEDDER__PROVIDER

Deployment Dependencies

Implementation Details

Key variables set by the template:

# Graphiti MCP
PORT=8000
SERVER__HOST=0.0.0.0
SERVER__TRANSPORT=http
DATABASE__PROVIDER=neo4j
NEO4J_URI=${{Neo4j.GRAPH_BOLT_URL}}
NEO4J_PASSWORD=${{Neo4j.GRAPH_DB_PASSWORD}}
OPENAI_API_KEY=
MODEL_NAME=gpt-4.1-mini
GRAPHITI_TELEMETRY_ENABLED=false

# Neo4j
NEO4J_AUTH=neo4j/${{secret(32)}}
NEO4J_server_default__listen__address=::
NEO4J_server_memory_heap_max__size=512m

First steps after deploy:

Open the MCP service and copy its public URL; the endpoint is https:///mcp (no trailing slash; /mcp/ redirects to an http URL).
Register it in your client, for example claude mcp add --transport http graphiti https:///mcp.
Ask your agent to remember something; it calls add_memory, and a few seconds later search_memory_facts returns the extracted facts.

Security note: the upstream MCP server has no authentication. Keep the URL private, or remove the public domain and connect from services inside the same Railway project through graphiti-mcp.railway.internal:8000, or front it with an authenticating proxy. Neo4j is never exposed publicly by this template; add a domain on port 7474 plus a TCP proxy on 7687 only if you want Neo4j Browser.

Why Deploy Graphiti MCP on Railway?

Railway is a singular platform to deploy your infrastructure stack. Railway will host your infrastructure so you don't have to deal with configuration, while allowing you to vertically and horizontally scale it.

By deploying Graphiti MCP on Railway, you get the official MCP server and a persistent Neo4j graph pre-wired over private networking, with generated secrets, health checks, and managed SSL, ready to plug into any MCP-capable agent.