Wanna do an MCP? Try this :-)

Community
1

MCP is a mechanism to expose things to AI agents.

Turns out it’s super easy to do, and it gives you some amazing leverage for tools! For example, I gave my kanban board a MCP server and now I can just say “move the note “whatever” to done” and it works ;-)

This post is to announce GitHub - ralsina/mcp: A crystal shard to create Model Context Protocol servers which is how I implemented that MCP as a spinoff shard.

It supports stdio MCP servers (nice way to expose a CLI tool to AI agents) and web MCP servers (with kemal at least)

Here is a FULL example of a MCP with a single tool, in stdio mode:

require "mcp"

# A simple tool that returns the answer to any question
class AnswerTool < MCP::AbstractTool
  @@tool_name = "get_answer"
  @@tool_description = "Returns 42 as the answer to any question you ask"
  @@tool_input_schema = {
    "type"       => "object",
    "properties" => {
      "question" => {
        "type"        => "string",
        "description" => "The question you want answered",
      },
    },
    "required" => ["question"],
  }.to_json

  def invoke(params : Hash(String, JSON::Any), env : HTTP::Server::Context? = nil)
    question = params["question"]?.try(&.as_s) || "unknown question"
    {
      "answer"   => 42,
      "question" => question,
    }
  end
end

# Start the stdio server - that's it! One line and you have a complete MCP server.
MCP::StdioHandler.start_server

Have fun and let me know if something is not ergnomic, I am trying to make this as boilerplate-free as possible.

6 Likes
2

When MCP servers started to become popular, I had Claude read the MCP specification and tried creating a Crystal library through vibe coding. I didn’t understand much then, and I still don’t now. But the code quickly grew beyond 10,000 lines, and I felt something wasn’t right, so I stopped. I think something smaller and more imperfect will eventually become widespread.

3

This implements most of the spec (didn’t implement resources because I don’t use them yet) and is 391 LOC (that includes the Kemal handler)

The spec is super simple.

It’s basically 4 entry points (initialize, list tools, invoke tool, resources) where the “complicated” one is invoke tool, which is jsonrpc, which is pretty simple.

2 Likes
4

I actually looked through the codebase and confirmed that the implementation is quite simple.
I still have some vague concerns about the future of MCP, but that’s another story.

5

Using this thread to also introduce another implementation of MCP protocol mcp.cr

1 Like
6

Really nice to see new members with real and interesting projects in Crystal.

The ToCry project is really nice, @ralsina!

1 Like
7

I have added a (fairly untested) implementation for resources and prompts so this now implements the whole MCP protocol.

1 Like
8

tried my hand at writing bindings a few months ago GitHub - nobodywasishere/mcprotocol: Crystal object bindings and serialization for the Model Context Protocol

9

Tried both shards, could not make them even build, decided to reimplement.

10

If you don’t mind, what was the issue you ran into?

11

I think in the case of your shard it was mostly that there are no docs or examples.

1 Like
12

so with the other one what was the issue you ran into?

13

I may use your shard when I rewrite my MCP server from Python to Crystal. It may happen soon.

1 Like
14

IIRC I got in a mess when specifying return types?

That was probably 99% my fault but was why I decided to go with “just return something that serializes to JSON”

1 Like