bhaskar511939/langgraph-boilerplate-kit

LangGraph Production Skeleton is a Python starter project for building AI chat agents that are meant to run reliably in a real company environment, not just in a prototype or tutorial. The author says the patterns come from internal AI infrastructure at a Fortune 5 company. The project addresses a common problem: LangGraph's own documentation examples show you how to build an AI agent, but those examples skip many engineering details that only matter at scale. This skeleton fills that gap. The core agent is built as a stateful conversation system, meaning it remembers the context of a conversation across multiple messages and across server restarts. In development you can keep conversation state in memory, in production you switch to MongoDB for durable storage that survives restarts and supports multiple servers running at the same time. Several additional features are included. A rolling summarization node watches how long conversations get and automatically compresses older messages into a summary when they exceed a threshold. This keeps AI inference costs from growing without limit. An MCP integration layer lets you connect external tool servers to the agent without rewriting the agent itself, so different teams can publish and update tools independently. All AI provider calls go through a single factory function that accepts any OpenAI-compatible API endpoint, so you can switch between OpenAI, Azure OpenAI, local models via Ollama, or others by changing two environment variables. The API layer uses FastAPI and streams responses to the browser using server-sent events, so users see output arriving word by word rather than waiting for the full response. The project includes a message sanitization layer that handles malformed conversation history, a known class of silent failures in production use. To get started, clone the repository, install dependencies, copy the example environment file, set an API key, and start the server. An interactive API documentation page is available at the local server address. License is MIT.