One endpoint. Your whole business.
Kula Intelligence is an MCP server. Point Claude, ChatGPT, Gemini or any MCP capable agent at it and you get the entire studio, members, bookings, payments, payroll, staff, classes and documents, as a single coherent queryable model. Not a dozen APIs stitched together in a prompt. One substrate, one contract, one connection.
One connection. Every client sees the same governed model of your business, with the same audit trail underneath.
Every tool has an API. None of them compose.
A member in the booking system, a customer in the payments system and a contact in the marketing system are three records with three ID schemes and no shared meaning. Hand that to an AI agent and it burns its context window guessing at joins, and guessing wrong. And composition is only where it starts.
Assembling what was actually asked often means chaining several calls, each returning a single record, repeated across the whole set. A question that should resolve in two seconds becomes a long sequence of requests before anything can be said at all.
Rate limits are set for jobs that run on a timetable with a known call count. Every integration on the account draws from the same allowance, so a question asked live can be throttled because something else ran first. Nothing has gone wrong. The budget was never meant for this.
Where two systems describe the same person there is no shared key and no agreed tie-break, because neither was built to be the other one’s source of truth. Somebody has to decide. Left alone, that somebody is the agent, mid-answer.
Per-call pricing is sensible for a nightly job with a predictable call count. An agent following a thread does not have a predictable call count, so the same rate card behaves very differently once questions are asked live.
None of this is a fault in anyone’s API. They were built to serve an application, on a schedule, one system at a time, and they do that well. It is simply not the shape a live question needs. That is a large part of why we built a model rather than a folder of connectors: the joins, the tie-breaks and the call budget are resolved once, in the substrate, before anyone asks anything. Our ontology documents cover a number of these cases directly, so an agent inherits the decision instead of working it out mid-answer.
Kula Intelligence does the composition work once, canonically, so agents never have to.
What sits behind the endpoint.
Four layers, all speaking the same identities, all reachable through the same endpoint. The graph and the event stream are live today. The document store and the CSV plane are next, and land behind the same contract.
The relationship graph
Members, staff and locations as typed nodes. Five edge types, with weights and decay. Who actually connects with whom, not just who transacted.
The event stream
Every action from every connected system, validated against one canonical envelope. A booking is a booking regardless of which platform it came from.
The document store
SOPs, plans and operating knowledge, normalised and retrievable into large context windows.
The CSV plane
Structured templates plus conversational ingest, so a system with no API, or no connector yet, can still reach the model.
Every field means exactly one thing.
The endpoint ships a machine readable semantic catalogue. When an agent asks which members are at risk, it is not pattern matching column names. It is reasoning against defined semantics: what an active member is, what an at risk signal means, how attendance and payment state relate.
Definitions travel with the data, so two agents asking the same question get the same answer.
This is what working at AI speed actually requires. Agents do not struggle because models are weak. They struggle because business data has no agreed meaning. We fixed the meaning.
Bring your own AI. Wire it in minutes.
Kula Intelligence works with the AI plan you already have. Add it as a connector in Claude, ChatGPT or Gemini, or wire it into your own stack: Claude Code, custom MCP clients, internal tooling. Your model choice, your inference costs, our substrate.
Add Kula Intelligence as a custom connector in your Claude settings, paste the endpoint, and authorise with your token. It is then available in every chat and project.
Create a token in the Kula console under Settings, Integrations, MCP Tokens, scoped to the role you want the agent to have. The console shows your region's exact endpoint. Treat the token like a database password.
No ETL. No warehouse project. No dashboard backlog.
Reads your systems directly.
Historical backfill on onboarding covers 18 to 24 months, so the model is useful immediately, not a year from now. Everything outside the live set is on the connector roadmap, and the CSV plane is coming to cover the rest: structured templates plus conversational ingest, so a studio on a system we have not connected yet does not have to wait for a bespoke build.
Frameworks in production right now.
The substrate is not theoretical. Two frameworks are live against real studio data, and both are examples of what anyone can build on the same endpoint.
Teacher bonus model
Instructor compensation calculated on retention and capacity together, where capacity is measured against the true median class size for that time slot, not how many bodies fit in the room. A 6am slot that realistically holds twelve is judged against twelve, not against a room built for twenty five. Compensation stops punishing teachers for the timetable and starts rewarding the thing that actually matters, which is whether members come back.
Yield analysis on fixed plans
Reads each new member’s first thirty days against their plan and surfaces the moments where the relationship needs a person, not a message. It runs independently of any outbound automation: the prompts go to the team to act on directly, rather than adding to the SMS and messaging sequence already running elsewhere. Automation handles the sequence. This handles the drift automation cannot see.
Instructor Bonus Builder
The teacher bonus model above, built as a real tool and released in the open. A deterministic engine splits a fixed pool on retention and capacity, ranked across the team and reconciled to the cent, and every allocation opens a Why panel that explains the number line by line. The AI explains the maths and pulls context. It never decides who gets paid what.
It reads live studio data through this endpoint, so it never touches GymMaster, Mindbody or Wix directly. Clone it, read how it queries the substrate, or lift the pattern for your own build.
Deterministic, hand-checked and golden-mastered in the repo's test suite. Same inputs, same settings, same result, every run.
Reads through safe views. Never raw access.
Every query, human or agent, first party or yours, passes through the same five layer policy engine. Every call declares a purpose. Every query is immutably logged: caller, purpose, entities queried, fields returned.
There is no relaxed external mode. An agent gets exactly what its key’s org and role permit, nothing more. Access is read only today. Write paths ship behind the same policy engine, with intent capture and authorisation flows, not before.
Your org, and only your org. Derived from the token, never from a parameter the caller sets.
What the key’s role is allowed to see, applied before anything is read.
Whose members, whose classes, whose locations this caller stands in relation to.
What the member has agreed to, honoured at query time rather than at export time.
Every call declares a purpose. Prohibited purposes are rejected regardless of scope.
The API problem isn’t a fitness problem.
Every industry has this shape. APIs were built for developers reading one system at a time, not for agents reasoning across all of them: paginated, rate limited, no shared identity, no agreed meaning. And even where an API is perfectly AI friendly, it still only hands back its own slice. Somebody still has to do the pulling together, and today that somebody is your agent, mid answer, with no idea it is guessing.
Kula Intelligence solves it for fitness and wellness by doing that composition once, in the substrate, where it can be governed and audited. The same substrate is built to take your business’s own sources and your own definitions: extend the semantic catalogue and what your business means by active, at risk or converted becomes what every agent reasons with. Extend it and you are not maintaining another integration. You are adding meaning to a model that already knows how to join.
The integration work happens once, or it happens in every prompt, forever.
The interfaces are yours. The substrate is constant.
Anything that speaks MCP can build on Kula Intelligence, and what you build can ship inside Kula Fitness so the team just finds it there. Every interface runs on the same governed model, with the same audit trail underneath. If you build tools or run services for studios, this is the substrate you build on instead of the integration work you would otherwise repeat.
The whole studio, answered in plain English inside Claude, ChatGPT or Gemini. No interface to build at all.
The joins are already done, so the front end is the only work left. Stand one up in an afternoon, not a quarter.
Target prospects who look like the members who actually stayed, using the retention signal rather than the demographic guess.
Every site on the same canonical model, so a group number means the same thing in every location.
Embed it inside Kula Fitness.
Anything built on the endpoint can run as a product inside Kula Fitness, the app your team is already in every day. You do not stand up accounts, invite anyone, or manage who can see what. That part is already solved, so the only thing you build is the thing itself.
Staff open it inside the app they already use. Nothing to roll out, nothing to explain, nothing for anyone to lose.
Access follows the roles you already maintain in Kula Fitness. Someone leaves, and their access to your tool leaves with them.
An embedded tool reads through the same policy engine and the same audit trail as everything else. Nothing gets a private door.
Which booking system do you run on?
Kula Intelligence reads it directly. Pick yours and we’ll take you to setup.