Connect External Services with MCP

Give your agent tools from any MCP server — GitHub, Notion, Postgres, and more — without writing integration code.

Step 1 — Connect a single server

from gantrygraph import GantryEngine
from gantrygraph.mcp import MCPClient
from langchain_anthropic import ChatAnthropic

agent = GantryEngine(
    llm=ChatAnthropic(model="claude-sonnet-4-6"),
    tools=[
        MCPClient("npx -y @modelcontextprotocol/server-filesystem /workspace"),
    ],
    max_steps=20,
)

result = agent.run("List every file modified today and show the 3 largest.")
print(result)

MCPClient launches the server subprocess, discovers its tools automatically, and terminates the process when the run ends — even on errors. GantryEngine manages the full lifecycle.

Step 2 — Connect multiple servers

from gantrygraph import GantryEngine
from gantrygraph.mcp import MCPClient, MCPToolRegistry
from langchain_anthropic import ChatAnthropic

registry = MCPToolRegistry([
    MCPClient("npx -y @modelcontextprotocol/server-filesystem /workspace"),
    MCPClient("npx -y @modelcontextprotocol/server-github"),
    MCPClient("npx -y @modelcontextprotocol/server-postgres postgresql://localhost/mydb"),
])

agent = GantryEngine(
    llm=ChatAnthropic(model="claude-sonnet-4-6"),
    tools=[registry],
    max_steps=30,
)

result = agent.run(
    "Find all open PRs tagged 'bug', check which ones have failing tests in "
    "the database, and write a report to /workspace/report.md."
)
print(result)

MCPToolRegistry opens all clients concurrently and presents their tools as a single flat list. Pass the registry as one item in tools=.

Step 3 — Mix MCP tools with custom tools

from gantrygraph import GantryEngine, gantry_tool
from gantrygraph.mcp import MCPClient
from gantrygraph.actions import FileSystemTools
from langchain_anthropic import ChatAnthropic

@gantry_tool
async def notify_slack(message: str) -> str:
    """Post an urgent notification to the #alerts Slack channel."""
    await slack_client.post("#alerts", message)
    return "posted"

agent = GantryEngine(
    llm=ChatAnthropic(model="claude-sonnet-4-6"),
    tools=[
        MCPClient("npx -y @modelcontextprotocol/server-github"),
        FileSystemTools(workspace="/workspace"),
        notify_slack,
    ],
    max_steps=20,
)

BaseAction instances, @gantry_tool functions, and MCPClient connectors all go in the same tools= list and are flattened into one registry.

---

Complete example

import asyncio
from gantrygraph import GantryEngine
from gantrygraph.mcp import MCPClient, MCPToolRegistry
from gantrygraph.memory import InMemoryStore
from langchain_anthropic import ChatAnthropic

llm = ChatAnthropic(model="claude-sonnet-4-6")

registry = MCPToolRegistry([
    MCPClient("npx -y @modelcontextprotocol/server-github"),
    MCPClient("npx -y @modelcontextprotocol/server-filesystem /workspace"),
])

agent = GantryEngine(
    llm=llm,
    tools=[registry],
    memory=InMemoryStore(),
    max_steps=40,
)

# Use arun() to avoid blocking in async contexts
result = asyncio.run(agent.arun(
    "Summarise the last 5 merged PRs in the repo, "
    "save the summary to /workspace/pr-digest.md, "
    "and return the filename."
))
print(result)

---

Variants

• Pass env vars to the server process: MCPClient("npx -y @mcp/github", env={"GITHUB_TOKEN": "ghp_..."})
• Use without an agent (inspect tools directly): call async with MCPClient(...) as c: print(c.get_tools())
• Preset shortcut: from gantrygraph.presets import mcp_agent; agent = mcp_agent(llm, "npx -y @mcp/github")

Troubleshooting

ImportError: mcp is a core dependency — run pip install gantrygraph (no extras needed, mcp ships with the base package).

Server subprocess fails to start — confirm npx is on your PATH and the package name is correct. Test standalone: npx -y @modelcontextprotocol/server-filesystem /tmp.

Tools not found after aenter — MCPClient.get_tools() returns an empty list when called before the client enters its context manager. Let GantryEngine manage the lifecycle by passing the client to tools= rather than entering it manually.

---

Next: Build custom tools · Run agents in parallel · Connect external services with MCP