library-mcp: working with Markdown knowledge bases

writing (41), llm (12), library-mcp (1)

At work, we’ve been building agentic workflows to support our internal Delivery team on various accounting, cash reconciliation, and operational tasks. To better guide that project, I wrote my own simple workflow tool as a learning project in January. Since then, the Model Context Protocol (MCP) has become a prominent solution for writing tools for agents, and I decided to spend some time writing an MCP server over the weekend to build a better intuition.

The output of that project is library-mcp, a simple MCP that you can use locally with tools like Claude Desktop to explore Markdown knowledge bases. I’m increasingly enamored with the idea of “datapacks” that I load into context windows with relevant work, and I am currently working to release my upcoming book in a “datapack” format that’s optimized for usage with LLMs. library-mcp allows any author to dynamically build datapacks relevant to their current question, as long as they have access to their content in Markdown files.

A few screenshots tell the story. First, here’s a list of the tools provided by this server. These tools give a variety of ways to search through content and pull that content into your context window.

The image displays a pop-up window titled “Available MCP tools” describing several server library tools like get_by_date_range and get_by_slug_or_url, which are used to retrieve posts or content based on specific parameters. The tools utilize Model Context Protocol (MCP) to interact with specialized servers.

Each time you access a tool for the first time in a chat, Claude Desktop prompts you to verify you want that tool to operate. This is a nice feature, and I think it’s particularly important that approval is done at the application layer, not at the agent layer. If agents approve their own usage, well, security is going to be quite a mess.

The image shows a prompt asking for permission to run a tool from a local library called “list_all_tags,” with options to allow the action for the chat, allow once, or deny. There is a warning about potentially malicious actions from MCP servers.

Here’s an example of retrieving all the tags to figure out what I’ve written about. You could do a follow-up like, “Get me posts I’ve written about ‘python’” after seeing the tags. The interesting thing here is you can combine retrieval and intelligence. For example, you could ask “Get me all the tags I’ve written, and find those that seem related to software architecture” and it does a good job of filtering.

The image shows a dialogue analyzing blog content to identify the most frequently covered topics, highlighting “Management” with 209 posts, followed by “Django” with 72 posts, and “Python” with 66 posts. The analysis also includes a list of less frequently covered topics such as innovation, plausible, tailscale, and business, each with one post.

Finally, here’s an example of actually using a datapack to answer a question. In this case, it’s evaluating how my writing has changed between 2018 and 2025.

The image shows a request to compare blog posts from January 2018 and January 2025, with a focus on how writing has evolved. The analysis highlights three main topics from the 2018 posts: technical infrastructure concepts, personal writing goals, and diversity in tech teams.

More practically, I’ve already experimented with friends writing their CTO onboarding plans with Your first 90 days as CTO as a datapack in the context window, and you can imagine the right datapacks allowing you to go much further. Writing a company policy with all the existing policies in a datapack, along with a document about how to write policies effectively, for example, would improve consistency and be likely to identify conflicting policies.

Altogether, I am currently enamored with the vision of useful datapacks facilitating creation, and hope that library-mcp is a useful tool for folks as we experiment our way towards this idea.