Implementing Functional LangGraph with MCP

ML Monday – Functional LangGraph with MCP

For the first Machine Learning Monday of this series, I spent a couple of hours setting up a basic LangGraph workflow that will be used to manage at least the main thread of an agentic workflow in my new side project. (See here for the daily plan).

I previously spent a weekend setting up a small python app demonstrating the use of Anthropic’s Model Context Protocol (MCP) that you can read about here. So, today’s work was mostly taking the parts I like from that, and deciding how to start splitting things up in a way that will allow me to take advantage of a serverless deployment architecture down the line.

This post is going to be a bit bare bones because I had to do a bunch of other stuff to get things going. In the future, expect pretty pictures/diagrams/examples, etc.

For now, here are a few thoughts/takeaways:

• I like using this dependency injection framework – It helps collect app configuration in one place, a config.yml file (or technically two, since you also need to set up the Containers that handle the dependency injection in containers.py). For testing, any part can be easily overridden via a with container.override(...): block.
• The new LangGraph Functional API seems pretty nice – It doesn’t end up saving LOC compared to the graph interface, but so far, I find it easier to read, and I hope that it will be easier to debug! You technically loose a bit of flexibility, but you can always call sub-graphs if needed to get back the deepest levels of control.
• MCP Client implementations are still pretty immature in both the official python-sdk and LangChain’s MCP adapters – I don’t want to re-invent the wheel, but I also want some insulation from breaking changes. I’m tackling this with a custom GraphRunAdapter where I exclusively implement methods that I’m actually interested in using internally, in addition to my own subclass of the LC MCP client implementation. The GraphRunAdapter uses my subclass of the LC client (so I can easily patch or fix behavior that I don’t like, without having to re-implement everything), and the rest of the code uses GraphRunAdapter so that any breaking changes from dependencies only need to be addressed in one place, and that’s the only place I ever need to think about the internals of langchain’s implementation.

In this session, I got to the point of the agent graph running with a test chat model, but I still have to deal with an issue related to a connection error between my MCP client and the “everything” MCP server.

Next steps

• Fix the MCP connection issue
• Check that tool calls can be made
• Check that the main agent can carry out multiple steps of thought before responding to the user

Further down the line

• Make the graph interruptable/resumable
• Make some configuration modifiable by user (e.g. prompt templates)
• Handle situations where context gets too large (either from long conversation, or tool returning too much information)
• Allow recursive call to the same graph as a sub-graph (this unlocks a LOT of customization of workflows via configuration only)