Model Context Protocol

Given Enough Context and Tools, A LLM can solve most problems.

• It can’t do everything, but it can a do a large number of useful things

• It can do things more effectively with more information

• It can use tools when it encounters scenarios that are solved sub optimally with token output eg: using code to do math instead of outputting a token that is probablistically next in a sequence.

How do you do this without reinventing the wheel every time a new model comes out?

• as an AI engineer, I don’t want to have to write a new way to call tools for each new model api that comes out.

• as a backend engineer, I don’t want to have to write a new service every time a new llm comes out to support its tool calling scheme.

Model Context Protocol is a USB for Model Context

• Lets standardize around a consensus on data transfer protocol between a llm and some data provider.

• Turns M ⨯ N integration problem by turning it into an M + N integration problem.

• Decouples AI client applications from AI tools and workflows for a platform.

How is it Implemented

• transport layer is http or by reading from the current process’ stdin and writing to stdout.

How do I write a mcp server?

from the offical python sdk readme

How do I use one

copied from my js code exections mcp server readme

Cursor

1. Install the server as above.

2. Create or edit .cursor/mcp.json in your project root:

{
  "mcpServers": {
    "js": {
      "command": "/usr/local/bin/mcp-v8",
      "args": [
          "--directory-path",
          "/tmp/mcp-v8-heaps",
      ]
    }
  }
}

1. Restart Cursor. The MCP tools will be available in the UI.

Demos

What are my open PRs based on the channel history of git? I’m robert wendt aka r33drichards

talking to kubernetes

port forward if not already

ssh -L 6443:localhost:6443 saphira.flakery.xyz   

whats are the resources on my cluster right now?

talk to a document with neo4j

port forward neo4j

ssh -L 7687:localhost:7687 44.229.154.130     

discussion on section schema and querying

Which led to this delightful query

MATCH (start:Section {clauseNumber: '10.6.4'}).                // match section with clauseNumber 10.6.4
OPTIONAL MATCH (start)-[:PARENT_OF*]->(parent:Section)         // recursively find its parents, if exists
OPTIONAL MATCH (start)-[:CHILD_OF*]->(child:Section)           // recursively find its children, if exists
OPTIONAL MATCH path=(start)-[:LINKS_TO*1..2]->(linked:Section) // recursively find its linked sections, 2 layers deep
WITH COLLECT(DISTINCT parent) AS parents, 
     COLLECT(DISTINCT child) AS children, 
     COLLECT(DISTINCT linked) AS linkedNodes, 
     start                                                     // collect these all to distinct values, and include start 
UNWIND parents + children + linkedNodes + [start] AS node      // flatten into one list
WITH DISTINCT node WHERE node IS NOT NULL                      // filter to distinct values again
RETURN node                                                    // return the node

what is my current schema? 

sierpinski’s trianlgle js exec

future

• llms have limited context windows, and some data is too large to fit in that.

• context is vast but pricey. prompting 51k tokens (1gb pdf file) costs $7.5k on google gen ai studio without even accouting for response tokens. They are also priced 4x more than input tokens.

• llms clients interact with mcp as a context provider, but 2nd generation of mcp will be helping models manipulate data in storage providers and manipulate that data using its metadata (schema)

• gen ui to interact with said data, without it flowing through the models context, to save on tokens.