tigergraph-mcp

Model Context Protocol (MCP) server for TigerGraph — lets AI agents interact with TigerGraph through the MCP standard. All tools use pyTigerGraph's async APIs for optimal performance.

Table of Contents

• Requirements

• Installation

• Getting Started

• Usage

  • Running the MCP Server

  • Configuration

  • Multiple Connection Profiles

  • Using with Existing Connection

• Client Examples

  • Using MultiServerMCPClient

  • Using MCP Client SDK Directly

• Available Tools

• LLM-Friendly Features

  • Structured Responses

  • Rich Tool Descriptions

  • Token Optimization

  • Tool Discovery

• Notes

Related MCP server: Apache AGE MCP Server

Requirements

• Python 3.10, 3.11, or 3.12

• TigerGraph 4.1 or later — Install from the TigerGraph Downloads page or use TigerGraph Savanna for a managed cloud instance.

Recommended: TigerGraph 4.2+ to enable TigerVector and advanced hybrid retrieval features.

Installation

Install with pip:

pip install tigergraph-mcp

Or with conda (from the tigergraph channel):

conda install -c tigergraph tigergraph-mcp

This installs:

• pyTigerGraph>=2.0.4 — the TigerGraph Python SDK

• mcp>=1.0.0 — the MCP SDK

• pydantic>=2.0.0 — for data validation

• click — for the CLI entry point

• python-dotenv>=1.0.0 — for loading .env files

To enable the tigergraph__generate_gsql and tigergraph__generate_cypher tools (LLM-powered query generation), install the optional [llm] extras (pip only):

pip install "tigergraph-mcp[llm]"

Getting Started

TigerGraph-MCP supports multiple AI agent frameworks. Choose the one that fits your workflow:

LangGraph (Recommended)

LangGraph is ideal for building stateful, agent-based workflows with complex tool chaining. Setup guide and full chatbot example:

• LangGraph Setup

• Chatbot example code

• Example transcripts

CrewAI

CrewAI provides a simpler starting point for basic agentic workflows with a web-based UI:

• CrewAI Setup

• Chatbot example code

GitHub Copilot Chat (VS Code)

For quick tasks or straightforward tool invocations directly in your editor:

• Copilot Chat Setup

Usage

Running the MCP Server

tigergraph-mcp

With a custom .env file:

tigergraph-mcp --env-file /path/to/.env

With verbose logging:

tigergraph-mcp -v    # INFO level
tigergraph-mcp -vv   # DEBUG level

Or programmatically:

from tigergraph_mcp import serve
import asyncio

asyncio.run(serve())

Configuration

The MCP server reads connection configuration from environment variables. You can set these either directly or in a .env file.

Using a .env File (Recommended)

Create a .env file in your project directory:

# .env — Username/Password authentication
TG_HOST=http://localhost
TG_GRAPHNAME=MyGraph  # Optional — can be omitted if the database has multiple graphs
TG_USERNAME=tigergraph
TG_PASSWORD=tigergraph
TG_RESTPP_PORT=9000
TG_GS_PORT=14240

Or use an API token instead of username/password:

# .env — API Token authentication
TG_HOST=http://localhost
TG_GRAPHNAME=MyGraph
TG_API_TOKEN=your_api_token_here

When TG_API_TOKEN (or TG_JWT_TOKEN) is set, the server uses token-based authentication (Authorization: Bearer <token>) and ignores username/password. You can obtain a token via pyTigerGraph's getToken() method or by directly calling TigerGraph's token generation endpoint.

When only username/password are provided and the TigerGraph instance requires a token for RESTPP endpoints, pyTigerGraph auto-mints one on the first 401 response and transparently retries the request — no manual token setup needed.

The server loads the .env file automatically. Environment variables take precedence over .env values.

Environment Variables

Multiple Connection Profiles

Define named profiles in your .env to work with multiple TigerGraph environments without changing any code.

Defining profiles

Each named profile uses a <PROFILE>_ prefix on the standard TG_* variables. Only variables that differ from the default need to be set.

# .env

# Default profile (no prefix) — password auth
TG_HOST=http://localhost
TG_USERNAME=tigergraph
TG_PASSWORD=tigergraph
TG_GRAPHNAME=MyGraph

# Staging profile — token auth
STAGING_TG_HOST=https://staging.example.com
STAGING_TG_API_TOKEN=staging_token_here
STAGING_TG_TGCLOUD=true

# Production profile — password auth
PROD_TG_HOST=https://prod.example.com
PROD_TG_USERNAME=admin
PROD_TG_PASSWORD=prod_secret
PROD_TG_GRAPHNAME=ProdGraph
PROD_TG_TGCLOUD=true

Profiles are discovered automatically at startup. Any variable matching <PROFILE>_TG_HOST registers a new profile. Values not set for a named profile fall back to the default profile's values.

Selecting the active profile

# Switch to staging for this run
TG_PROFILE=staging tigergraph-mcp

# Or set permanently in .env
TG_PROFILE=prod

If TG_PROFILE is not set, the default profile is used.

Switching profiles per call

Every tool accepts an optional profile argument, so an agent can route individual calls to different environments without restarting the server. Connections are pooled per profile and reused across calls.

User: Compare the vertex count of MyGraph between staging and prod.

Agent:
  → get_vertex_count(profile="staging", graph_name="MyGraph")
  → get_vertex_count(profile="prod",    graph_name="MyGraph")

User: Show me the schema on staging, then run this GSQL on prod:
      SHOW VERTEX Person

Agent:
  → get_graph_schema(profile="staging", graph_name="MyGraph")
  → gsql(profile="prod", command="SHOW VERTEX Person")

A minimal system prompt that lets the agent discover profiles at runtime:

You are a TigerGraph assistant with access to multiple environments
through the tigergraph-mcp server.

At the start of a session — or whenever the user references an
environment you haven't seen yet — call the `list_connections` tool
to discover available profiles. Do not assume or hardcode profile names.

Most tools accept an optional `profile` argument. When the user names
an environment, pass `profile="<name>"` to the tool calls in that turn.
If the user doesn't specify a profile, use the default profile.

Omitting profile falls back to TG_PROFILE, then "default".

Using with Existing Connection

from pyTigerGraph import AsyncTigerGraphConnection
from tigergraph_mcp import ConnectionManager

async with AsyncTigerGraphConnection(
    host="http://localhost",
    graphname="MyGraph",
    username="tigergraph",
    password="tigergraph",
) as conn:
    ConnectionManager.set_default_connection(conn)
    # ... run MCP tools ...
# HTTP connection pool is released on exit

Client Examples

Using MultiServerMCPClient

from langchain_mcp_adapters import MultiServerMCPClient
from pathlib import Path
from dotenv import dotenv_values
import asyncio

env_dict = dotenv_values(dotenv_path=Path(".env").expanduser().resolve())

client = MultiServerMCPClient(
    {
        "tigergraph-mcp-server": {
            "transport": "stdio",
            "command": "tigergraph-mcp",
            "args": ["-vv"],
            "env": env_dict,
        },
    }
)

tools = asyncio.run(client.get_tools())

Note: You can pass your TigerGraph database credentials directly into the "env" mapping:

    "env": {
      "TG_HOST": "http://localhost",
      "TG_USERNAME": "tigergraph",
      "TG_PASSWORD": "tigergraph",
      "TG_GRAPHNAME": "MyGraph"
    }

Using MCP Client SDK Directly

import asyncio
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

async def call_tool():
    server_params = StdioServerParameters(
        command="tigergraph-mcp",
        args=["-vv"],
    )

    async with stdio_client(server_params) as (read, write):
        async with ClientSession(read, write) as session:
            await session.initialize()

            tools = await session.list_tools()
            print(f"Available tools: {[t.name for t in tools.tools]}")

            result = await session.call_tool(
                "tigergraph__list_graphs",
                arguments={}
            )
            for content in result.content:
                print(content.text)

asyncio.run(call_tool())

Available Tools

Global Schema Operations

• tigergraph__get_global_schema — Get the complete global schema via GSQL LS

Graph Operations

• tigergraph__list_graphs — List all graph names in the database

• tigergraph__create_graph — Create a new graph with schema

• tigergraph__drop_graph — Drop a graph and its schema

• tigergraph__clear_graph_data — Clear all data from a graph (keeps schema)

Schema Operations

• tigergraph__get_graph_schema — Get schema as structured JSON

• tigergraph__show_graph_details — Show schema, queries, loading jobs, and data sources

Node Operations

• tigergraph__add_node / tigergraph__add_nodes

• tigergraph__get_node / tigergraph__get_nodes

• tigergraph__delete_node / tigergraph__delete_nodes

• tigergraph__has_node

• tigergraph__get_node_edges

Edge Operations

• tigergraph__add_edge / tigergraph__add_edges

• tigergraph__get_edge / tigergraph__get_edges

• tigergraph__delete_edge / tigergraph__delete_edges

• tigergraph__has_edge

Query Operations

• tigergraph__run_query — Run an interpreted query

• tigergraph__run_installed_query — Run an installed query

• tigergraph__install_query / tigergraph__drop_query

• tigergraph__show_query / tigergraph__get_query_metadata / tigergraph__is_query_installed

• tigergraph__get_neighbors

Loading Job Operations

• tigergraph__create_loading_job

• tigergraph__run_loading_job_with_file / tigergraph__run_loading_job_with_data

• tigergraph__get_loading_jobs / tigergraph__get_loading_job_status

• tigergraph__drop_loading_job

Statistics Operations

• tigergraph__get_vertex_count / tigergraph__get_edge_count

• tigergraph__get_node_degree

GSQL Operations

• tigergraph__gsql — Execute raw GSQL

• tigergraph__generate_gsql — Generate GSQL from natural language (requires [llm])

• tigergraph__generate_cypher — Generate openCypher from natural language (requires [llm])

Vector Schema Operations

• tigergraph__add_vector_attribute / tigergraph__drop_vector_attribute

• tigergraph__list_vector_attributes / tigergraph__get_vector_index_status

Vector Data Operations

• tigergraph__upsert_vectors

• tigergraph__load_vectors_from_csv / tigergraph__load_vectors_from_json

• tigergraph__search_top_k_similarity / tigergraph__fetch_vector

Data Source Operations

• tigergraph__create_data_source / tigergraph__update_data_source

• tigergraph__get_data_source / tigergraph__drop_data_source

• tigergraph__get_all_data_sources / tigergraph__drop_all_data_sources

• tigergraph__preview_sample_data

Discovery & Navigation

• tigergraph__discover_tools — Search for tools by description or keywords

• tigergraph__get_workflow — Get step-by-step workflow templates

• tigergraph__get_tool_info — Get detailed information about a specific tool

LLM-Friendly Features

Structured Responses

Every tool returns a consistent JSON structure:

{
  "success": true,
  "operation": "get_node",
  "summary": "Found vertex 'p123' of type 'Person'",
  "data": { ... },
  "suggestions": ["View connected edges: get_node_edges(...)"],
  "metadata": { "graph_name": "MyGraph" }
}

Error responses include actionable recovery hints:

{
  "success": false,
  "operation": "get_node",
  "error": "Vertex not found",
  "suggestions": ["Verify the vertex_id is correct"]
}

Rich Tool Descriptions

Each tool includes detailed descriptions with use cases, common workflows, tips, warnings, and related tools.

Token Optimization

Responses are designed for efficient LLM token usage — no echoing of input parameters, only new information (results, counts, boolean answers).

Tool Discovery

# Find the right tool
result = await session.call_tool("tigergraph__discover_tools",
    arguments={"query": "how to add data to the graph"})

# Get a workflow template
result = await session.call_tool("tigergraph__get_workflow",
    arguments={"workflow_type": "data_loading"})

# Get detailed tool info
result = await session.call_tool("tigergraph__get_tool_info",
    arguments={"tool_name": "tigergraph__add_node"})

Notes

• Transport: stdio by default

• Error Detection: GSQL operations include error detection for syntax and semantic errors

• Connection Management: Connections are pooled by profile and reused across requests; pool is released at server shutdown

• Performance: Persistent HTTP connection pool per profile; async non-blocking I/O; v.outdegree() for O(1) degree counting; batch operations for multiple vertices/edges