AI assistants are good at recommending things and bad at proving the recommendation was real. Studio Referrals is a small MCP server that closes part of that gap for fitness and yoga classes, starting with independent studios in Los Angeles.

What it does. Five tools over streamable HTTP at https://studio-referrals.t6y11a.workers.dev/mcp:

  • list_studios and search_offerings return real studio offerings compiled from the studios' own public websites, with the source URL and a checked_at timestamp.
  • Every studio carries an inventory block: when we last verified the published schedule and the next dated, bookable sessions (for our first Venice studio: three $30 classes the week of Dec 2, verified on her booking page, explicitly labeled as future dates, never "this week").
  • get_booking_channel returns the studio's own booking route. We never book on the studio's systems; the consumer books and pays the studio directly.
  • create_referral records an attributed lead (referral ID + first name only) so a studio can recognize a referred client at settlement; get_referral lets the referring agent confirm the record and never returns the name.

What it deliberately is not. It is not a booking engine, not a payment rail, and not a scraped listing farm. Studios are listed from public sources and a listing is not a partnership; referral fees are agreed with a studio separately before anything is billed. Anonymous page click-throughs are counted as leads only, never billed.

Why the date handling is strict. The failure mode of directory-style MCP servers is stale claims. A consumer asking for "yoga this week" must not be sold a class three weeks out. So dated sessions carry explicit dates and a machine-readable date_guidance field instructing consuming agents not to present future sessions as current-week availability.

Listed and ownership-verified on agent-tools.cloud. Feedback welcome - it's a pilot, and the honest mechanics matter more to us than the numbers.