Okay so I wasn't planning on writing about AI stuff again this month, I did that back in the summer and told myself I'd cool it. But I stayed up way later than I meant to on Sunday messing with something Anthropic put out last week, and I can't not write about it, so here we are.
The thing is called MCP, Model Context Protocol, and Anthropic open-sourced it on the 25th. The pitch is pretty simple once you strip out the announcement-blog-post language: it's a standard way for an AI assistant to talk to stuff outside the chat window. Your files, a database, Slack, GitHub, whatever. Before this, every single one of those connections was some bespoke plugin built by whoever felt like building it, and none of them worked the same way or lived in the same place. MCP is basically Anthropic saying, here's one protocol, write a server for it once, and any compatible client can use it.
I know, I know. Zapier has been selling some version of "connect your apps" since 2011 (which, funny enough, is the same year I started this blog). IFTTT too. Every dev tools startup pitch deck for the last five years has had a slide that says some version of "the future is agents that can take action in your tools." So on paper this isn't a new idea. What's different is it's coming from the company that makes the model, it's free, and it's not tied to a specific SaaS product trying to lock you in.
What I actually did
Claude Desktop already supports it, so I didn't need to build anything to test it, just configure a server. There's a config file, claude_desktop_config.json, that you have to hand-edit, which already tells you this is very much a v1 developer-facing thing and not something my mom is going to be doing anytime soon. I pointed a filesystem server at a folder where I keep old blog drafts and half-finished notes going back years, some of which I genuinely forgot existed.
Had to restart the app three separate times before it picked up the config correctly. First time I had a trailing comma in the JSON (rookie mistake, I know better), second time I just didn't fully quit the app, I minimized it, which on a Mac doesn't count. Third time it worked.
And then I asked Claude what was in the folder and it just... told me. Correctly. It found a draft from early 2020 that I'd apparently titled "why does everyone hate open plan offices (draft, unfinished, do not post??)" which, fair, I never did post that one and I stand by it staying unpublished.
It's a small thing but it felt different from copy-pasting text into a chat box, which is what I've been doing for over a year now whenever I want an AI to look at something of mine. Here the assistant is just reaching into the folder itself.
The part that bugs me
Setup friction aside, my actual complaint is that this only feels useful to people who are already comfortable in a terminal and a JSON file. I don't think that's a knock on Anthropic specifically, most "open protocol, here's the spec, build your own server" launches start exactly like this. But it does mean the gap between "power user with a filesystem server running on a Sunday night" and "normal person who just wants their AI to see their stuff" is still enormous, and somebody is going to have to build the pretty wrapper around this before it means anything to anyone outside our little corner of the internet.
There are already a handful of pre-built servers floating around for things like Google Drive, Postgres, and GitHub, and I'd bet by next year there's a small ecosystem of these things, some official, most not. Whether that ecosystem actually gets used or just sits there like half the VS Code extension marketplace, I have no idea yet. Ask me again in six months.
Anyway. Thanksgiving's Thursday, I've got a turkey to not screw up, and apparently a decade-old blog draft to maybe finally finish.