@nestjs-agentic/langgraph

Experimental LangChain/LangGraph compatibility package for the NestJS-native runtime for governed AI agents.

Status and Compatibility Scope

Experimental: this adapter is intended for evaluation of type compatibility and governed tool wrapping. It is not a complete LangGraph agent runtime.

The current implementation:

• wraps ResolvedTool closures as LangChain tools;
• accepts a LangChain model and LangGraph BaseCheckpointSaver types;
• does not build, compile, or execute a LangGraph StateGraph;
• invokes a configured model once and does not execute returned tool calls or run a model/tool loop;
• uses a synthetic fallback when no model is configured, directly invoking tools with generated test arguments;
• exposes synthetic tool events from stream() rather than model or graph token streaming; and
• does not provide unified durable recovery through core execution state.

The fallback and stream paths can invoke application tools with generated arguments that do not match the user request or tool schema. Do not use this adapter with production side effects.

Installation

npm install @nestjs-agentic/langgraph @langchain/core @langchain/langgraph

Model Compatibility Path

Pass a configured instance explicitly to select the current single-invoke model path:

import { Module } from '@nestjs/common';
import { MemorySaver } from '@langchain/langgraph';
import { AgenticModule, RUNTIME_ADAPTER } from 'nestjs-agentic';
import { LangGraphRuntimeAdapter } from '@nestjs-agentic/langgraph';

@Module({
  imports: [
    AgenticModule.forRoot({
      defaultModel: { provider: 'custom', model: 'configured-chat-model' },
    }),
  ],
  providers: [
    {
      provide: RUNTIME_ADAPTER,
      useFactory: () => new LangGraphRuntimeAdapter({
        model,
        checkpointer: new MemorySaver(),
      }),
    },
  ],
})
export class AppModule {}

This registration binds tools to one model invocation. It does not add graph nodes, conditional edges, tool-call execution, or a second model round.

Checkpointer Scope

When no model is supplied, the fallback path can write a synthetic checkpoint through the configured saver. This is not equivalent to a compiled graph checkpoint or durable AgentRunner recovery. The model path and stream() do not currently provide the same checkpoint behavior.

Use MockRuntimeAdapter for deterministic tool and policy tests. See the product roadmap for planned common runtime semantics.

License

MIT © irzix