LangChain and LangGraph integration

Published boundary: @theorvane/type-mcp/langchain is part of the published @theorvane/type-mcp@0.2.2 package. It is tools-only; LangGraph remains a consumer-owned composition choice.

Scope

@theorvane/type-mcp/langchain converts decorated @McpTool methods into LangChain DynamicStructuredTool instances. It is tools only:

included: tool name, description, Zod input schema, explicit InstanceResolver, safe invocation results;
excluded: MCP server startup, HTTP/stdio transport, resources, prompts, models, agent executors, graph topology, persistence, and authorization policy.

The core package and @theorvane/type-mcp/http remain independent of agent frameworks and graph runtimes; only the isolated @theorvane/type-mcp/langchain subpath depends on the LangChain peer.

Dependency boundary

The adapter has an optional peer dependency on @langchain/core. A consumer that imports the adapter must install a compatible peer:

npm install @theorvane/type-mcp@0.2.2 @langchain/core zod

LangGraph is a consumer choice, not an adapter dependency. Install it only when using a graph:

npm install @langchain/langgraph

Convert decorated tools

import { createLangChainTools } from "@theorvane/type-mcp/langchain";

const tools = await createLangChainTools(CatalogTools);
const result = await tools[0].invoke({ sku: "SKU-7" });

The adapter reads existing TypeMCP metadata, validates it, resolves the server once, and creates one LangChain structured tool per decorated MCP tool. Zod rejects invalid input before application code runs. String handler results are returned unchanged; JSON-compatible non-strings are returned as JSON text. Missing handlers, rejected handlers, cyclic values, and undefined results resolve to Tool execution failed without exposing application errors.

For a dependency-requiring server, use the existing explicit resolver seam:

const tools = await createLangChainTools(InjectedTools, {
  resolver: { resolve: () => new InjectedTools(catalogService) },
});

LangGraph composition

The consumer owns its graph and policies. Pass a mutable copy of the adapter result to LangGraph's ToolNode:

import { ToolNode } from "@langchain/langgraph/prebuilt";
import { createLangChainTools } from "@theorvane/type-mcp/langchain";

const tools = await createLangChainTools(CatalogTools);
const toolNode = new ToolNode([...tools]);

TypeMCP does not create a graph, select a model, route messages, store state, approve calls, or host a service. Add toolNode to the consumer's StateGraph and own those decisions there.

Repository example and verification

The LangGraph ToolNode example (opens in a new tab) is an in-memory repository-source example. It has no model, API key, network listener, persistence store, or external request.

npm ci
npm run example:langgraph-tools
npm test -- --run test/langgraph-tool-node.test.ts

View source at 9124c25519f3 (opens in a new tab)