Connecting Claude to your LUVI account: what the 35 MCP tools actually do

LUVI runs a Model Context Protocol server, which means an assistant like Claude or ChatGPT can act on your LUVI account directly rather than telling you what to click. The live connector documentation lists 35 tools. That number is less interesting than their shape: they are grouped so that an assistant has to look before it spends, and the guardrails are in the tools rather than in a warning.

What can it actually do?

Five things, in the order a sensible assistant does them.

• Look at the catalog. List the models, and fetch one model's exact contract—every parameter, its enum values with their exact casing, its defaults, and the reference slots with their keys and limits. This is the tool that stops an assistant guessing that a resolution is written 2K when the model only accepts 2k.
• Read the manual. Fetch the prompting guide for the family, then run the free check over a finished prompt. Same two layers the feather button runs in the Workspace, exposed as tools.
• Quote before spending. Estimate the cost of an exact configuration in credits, so the assistant can tell you what a run will cost before it asks to make it.
• Generate and follow it. Start a generation, poll its status, cancel one that is still queued.
• Report what happened. Every finished generation carries the credits it actually charged, and a transactions tool returns the same itemized ledger you see in your account.

There is also a group for projects and folders, one for uploads, and eight tools for building and running workflows. One more searches the Prompt Lab styles and the Pose Atlas body poses, so an assistant puts a real style or pose on a workflow node instead of guessing one.

Why does it report the real charge?

Because an estimate is a promise and a ledger is a fact, and for a long time an assistant working remotely could tell you what it had made but not what it had cost. The generation-status, asset and library tools now carry the credits actually debited—the real ledger figure, not a re-quote—and the transactions tool returns the itemized list.

Two things are worth knowing about that number. It is denominated in credits and never in currency, exactly as your account page shows it. And it can lag: a generation is marked complete a moment before its ledger row settles, so a very fast poll can find no charge recorded yet. A missing row means "not settled yet", not "free".

What can it not do?

The limits are deliberate, and two of them surprise people.

It cannot take a file from your chat. The protocol passes tool arguments as strict JSON, so the bytes of an image you attached to a conversation cannot reach a tool. The connector solves this in two ways instead: a tool that hands you a deep link opening the upload dialog already locked to the right project and folder, and a ticket flow that lets an assistant with code execution push the bytes straight to storage and then confirm them. Neither one pretends to smuggle the file through the conversation.

It cannot ask you to confirm a spend on every client. Dialog prompts of that kind are an optional part of the protocol, and not every client implements them. The connector therefore does not rely on one existing—which is why irreversible operations are built to be undoable instead. Deleting a folder, for example, does not delete anything immediately: the folder leaves your lists, its contents stay where they are, and a scheduled job finishes the job a day later. Until then it can be restored.

Where does the queue come in?

An account can have a limited number of generations in flight at once. Rather than returning an error when an assistant hits that ceiling, the connector accepts the request and queues it: you get a generation id straight away with a queued status, and the work starts when a slot frees. Nothing is charged while a job waits, the order is preserved, and a queued job can still be cancelled. Once it has started, it cannot—there is no cancel upstream.

How we checked

• What was read: the live connector documentation page, and the tool surface itself as exercised against a real account—catalog listing, model contracts, prompting manuals, prompt checks, cost estimates, folder creation, generation and status polling.
• Date: September 22, 2026. The tool count comes from the live documentation page title on that date.
• Results: 34 tools on that date; on October 20, 2026 the page title says 35, after the style and pose search was added. Model contracts returned exact enum values and case; estimates returned credits; finished generations returned the credits actually charged.
• Drawbacks: This post describes the tool surface, not the quality of any assistant's judgment when using it. Tool counts change as the product changes—the documentation page is the live figure.

In LUVI

You add the connector in your assistant's own settings as a custom connector, then authorize it against your LUVI account. From then on the assistant is acting as you: it spends your credits, writes into your projects, and appears in your library alongside everything you made yourself.

The order that works best mirrors the one in the Workspace. Ask it to fetch the model's contract and the family's manual first, have it check the prompt, look at the estimate, and only then let it generate. All of that except the generation is free.

The full tool list, with what each one takes and returns, is on the connector documentation page.

New here? Create a LUVI account, connect it, and start by asking your assistant what a specific setting would cost. It is the cheapest way to see whether the two of you agree about the plan.

Sources

• Model Context Protocol — Model Context Protocol, specification site. Accessed September 22, 2026.