Connecting Claude to My X Bookmarks with MCP: A Walkthrough

I built the MCP tools into Birdfeeder a few weeks ago and kept meaning to actually use them myself instead of just shipping the feature and moving on. Today I sat down, generated a token, and wired it into Claude Desktop to see what it's actually like from the other side.

The setup itself is small. Birdfeeder already had a personal access token on the user model, the same one that used to power a browser extension. I didn't want a second auth system just for this, so the MCP server checks that same token as a Bearer header. One less thing to explain, one less thing to break.

In Settings there's now a token field. I copied mine, opened Claude Desktop's config file, and added an entry pointing at Birdfeeder's MCP endpoint with the token in the Authorization header. Restarted Claude, and a new tool showed up in the sidebar. No OAuth dance, no separate account, just paste and go.

First thing I asked was something dumb, just to see if it worked: "What have I bookmarked about Rust error handling?" Claude called the search tool, got back three bookmarks, and gave me a plain-language answer with the tweets quoted and linked. It read like it had actually looked, not like it was guessing from a cached summary.

Then I asked which accounts I bookmark the most from, expecting maybe a novelty answer. It used the second tool, the one that aggregates by account, and came back with a list that was basically correct going by memory. That one surprised me a bit, because I'd forgotten I built it to be genuinely useful and not just a demo tool.

One thing to know if you try this on a non-subscriber account: search through Claude has the same 7-day lookback that the website search has. I didn't special-case MCP to bypass entitlement checks, so if you're on the free trial and ask about something you bookmarked three weeks ago, Claude will come back with nothing, same as if you'd typed the query into the website yourself. That's intentional. It would be a weird kind of bug if the AI assistant had access to more of your own data than the website did.

Under the hood the MCP server is its own process, not bundled into the Rails app that serves getbirdfeeder.com. I did that on purpose. An AI assistant poking around your bookmarks is a different traffic pattern than someone loading a page, and I didn't want a slow tool call from Claude to have any chance of affecting page load times for everyone else. It runs on its own port, gets its own pidfile, and if it falls over it doesn't take the website with it.

It's also not Claude-specific, even though that's what I tested with today. MCP is a protocol Anthropic published, not something proprietary to their apps, so any MCP-compatible assistant can point at the same endpoint with the same token and get the same two tools.

Nothing about this changes what Birdfeeder does at its core. It's still $5 a month or $45 a year with a 14-day trial and no card required up front. It just means if you'd rather ask a question in plain English than scroll through search results, you can now do that from wherever you already keep Claude open.