We shipped a 12-tool MCP connector for marketers. Here's what building it actually looked like.
Six platforms. Twelve tools. One connector URL. Here's how we built it.
When Anthropic launched MCP (Model Context Protocol), the opportunity was clear. Claude could stop being a document reader and start being an actual data agent. Give it the right connector, and it can fetch, structure, and reason over live data, all inside one conversation.
The use case we kept coming back to was marketing research. Marketers live inside Claude for strategy and copy, but they always had to prep data themselves first. We wanted to close that gap.
The hardest part wasn't building the tools. It was the author.
Getting OAuth to work cleanly inside a Claude conversation, where a user clicks a link, logs in, and returns to a working connector, took far longer than expected. The UX has to be completely frictionless, or people drop off. We ended up using Supabase Auth with short-lived tokens,s so Claude never sees a password. That also handled the security question cleanly.
The pricing model was the other big decision. Subscriptions didn't fit. Marketers have spiky usage: quiet weeks, then a campaign sprint where they're pulling hundreds of data points. Credit packs matched that pattern better. And credits never expire, so there's no pressure to burn through them.
The biggest surprise was the Google Maps use case. We built it expecting agencies to use it for competitor research. The most enthusiastic early users turned out to be brand managers auditing their own reviews. "Get the top 100 reviews for our locations and tell me what customers complain about most" turns out to be an incredibly high-value task that used to eat a full analyst afternoon. Set up for the connector is: copy URL, paste into Claude settings, sign in. Under 60 seconds. Works across Claude.ai, Claude Desktop, Claude Code, and Cowork.
We're at 12 tools across Instagram, TikTok, YouTube, Google Maps, Meta Ads Library, and Google Keyword Planner. More in progress.
If you're building on MCP, we're happy to talk through the auth implementation. It's the part with the least documentation available.
Great write-up! The OAuth friction you mentioned is definitely a hidden hurdle in MCP adoption—glad to see you found a smooth path with Supabase. I also really like the 'credit pack' approach; spiky usage is so common in agent-driven tasks. For my project, AgentShare (focused on price & offer infrastructure), we're seeing similar patterns. Making data truly 'agent-ready' via MCP is a game changer for marketers. Quick question: How are you handling token limits when pulling large amounts of review data back into the context window?