Connecting Letta to Local MCP Servers

stdio transport launches MCP servers as local subprocesses, ideal for development and testing. Local (stdio) MCP servers can be useful for local development, testing, and situations where the MCP server you want to use is only available via stdio.

Setup

ADE: Tool Manager → Add MCP Server → stdio → specify command and args

1
import { LettaClient } from '@letta-ai/letta-client'
2
3
// Self-hosted only
4
const client = new LettaClient({
5
    baseUrl: "http://localhost:8283"
6
});
7
8
// Connect a stdio server (npx example - works in Docker!)
9
const stdioConfig = {
10
    server_name: "github-server",
11
    command: "npx",
12
    args: ["-y", "@modelcontextprotocol/server-github"],
13
    env: {"GITHUB_PERSONAL_ACCESS_TOKEN": "your-token"}
14
};
15
16
await client.tools.addMcpServer(stdioConfig);
17
18
// List available tools
19
const tools = await client.tools.listMcpToolsByServer("github-server");
20
21
// Add a tool to use with agents
22
const tool = await client.tools.addMcpTool("github-server", "create_repository");

Docker Support

Letta’s Docker image includes npx, so npm-based MCP servers work out of the box. Custom Python scripts or missing dependencies require workarounds.

• Works in Docker: npx servers from the official MCP repository
• Challenging: Custom scripts, local file paths, missing system dependencies
• Alternatives: Use remote servers or mcp-proxy

Troubleshooting

• Server won’t start: Check command path, dependencies, environment variables
• Connection fails: Review Letta logs, test command manually
• Tools missing: Verify MCP protocol implementation and tool registration