Vi3W
Skip to content

Guides

Do MCP tool calls use credits or tokens? What a 3D session really costs

MCP tool calls cost assistant tokens and sometimes service credits. How both add up in a 3D session with Claude or ChatGPT, and how to spend less.

5 min read

A glass jar of glowing violet tokens with more tokens floating out of it
Illustration. A 3D chat session draws from two jars: the assistant's and the tool's.

Quick answer

Do MCP tool calls cost tokens or credits?

Usually both, from two separate budgets. The assistant spends tokens, or plan usage, reading tool lists, your messages and every tool result. The connected service may also charge its own credits for real work, such as generating, rigging or retexturing a 3D model. The protocol itself is free.

Two meters run at once

When you ask Claude or ChatGPT to make a 3D model through an MCP server, two different things get used:

  1. The assistant's budget. On the API that's tokens you pay for. On a Claude or ChatGPT subscription it's your plan's usage allowance. Every word the model reads or writes counts.
  2. The tool's budget. If the server wraps a paid service, that service charges in its own currency (credits, XP, or money) for the actual work.

MCP itself, the protocol, costs nothing. It's an open standard. The costs come from the two ends it connects. If tokens and credits still feel like the same thing, AI credits vs tokens explains the difference in plain terms.

Where the assistant's tokens go

A tool call isn't free for the model just because the work happens elsewhere. Here's what passes through its context in a typical session:

  • Tool definitions. Every connected server sends a list of tools with names, descriptions and input schemas. The model reads all of them so it can decide what to call. Ten connected servers means ten lists, on every turn.
  • Your messages. The prompt you type, plus everything earlier in the conversation.
  • The model's tool calls. The arguments it writes ("prompt: a mossy stone well, stylised, 1.5 m tall").
  • Tool results. Whatever the server sends back. A short result ("model ready, link here") is cheap. A full Blender scene summary with 400 objects, or a viewport screenshot, is not.
  • The model's reply explaining what it did.

This is why a long Blender MCP session can eat a plan's allowance faster than a long chat: every screenshot and scene dump is read in full, and the conversation keeps growing.

Where the tool's credits go

The second meter belongs to the service. With a 3D generation service, the expensive part is GPU time, so that's what's charged. In Vi3W, generation and the 3D tools (Add bones, Make it move, New paint job and Clean shape) each have a cost in XP that's shown before the job runs, and that's true whether you click the button in the app or an assistant calls it through Vi3W's MCP server. Conversion to OBJ, STL, PLY or glTF and Make it lighter are free. How XP and credits work on the plan side is in AI credits explained.

Two things are worth knowing:

  • The assistant doesn't pay the service. Your Claude or ChatGPT subscription doesn't cover Vi3W's XP, and Vi3W's plan doesn't cover Claude's usage. They're separate accounts.
  • A retry costs twice. If the model calls a generation tool, doesn't like the result and calls it again, that's two jobs on the service side. Telling the assistant "ask me before you regenerate" keeps that under your control.
A glass chat bubble linked by a bridge of light to a 3D scene
Illustration. Each round trip costs the assistant tokens; heavy work may also cost the service's credits.

A worked example: one prop, start to finish

Here's what happens when you ask Claude, with Vi3W connected, for a game-ready lantern. No numbers, because prices change; just where each cost lands.

  1. You type: "Make a stylised brass lantern for a fantasy game, about 40 cm tall, then give me an FBX." Assistant tokens: your message plus the tool lists of every connected server.
  2. Claude calls Vi3W's text-to-3D tool. Assistant tokens: a short tool call. Service cost: one generation, at the XP price shown.
  3. The result comes back. Assistant tokens: the result text.
  4. You say: "Make the metal look older and greener." Claude calls the retexture tool (New paint job). Service cost: one retexture.
  5. You ask for a lighter mesh. Claude calls Clean shape at a lower model size. Service cost: one Clean shape run.
  6. You download the FBX. Conversion is free.

Five turns, three paid jobs on the service side, and a modest amount of assistant usage, because the results were short. The same session in Blender MCP, with a screenshot after every step, would use far more assistant tokens and no service credits at all.

How to spend less on both sides

On the assistant side

  • Connect only what you need. Disable connectors you aren't using in this chat. Fewer tool lists, fewer tokens on every turn.
  • Ask for summaries. "Tell me the object count and the five heaviest meshes" instead of "list everything in the scene".
  • Use screenshots sparingly. Ask for one when you need to see the layout, not after every move.
  • Start fresh. When a session has gone long and wandered, start a new chat with a two-line summary of where you are. The old context stops being re-read.

On the service side

  • Get the prompt right before generating. Agree the description, style and size with the assistant in text first, then make one generation call. Tips are in prompting a chat for 3D models.
  • Ask for confirmation before paid calls. Put it in your first message: "Before any generation, rig, retexture or remesh, tell me what you're about to run and wait for my OK."
  • Batch with intent. For a set of props, write the list first, review it, then generate. Generating a batch of game props from one chat shows the workflow.
  • Use free steps where they fit. Format conversion and Make it lighter don't cost XP.

What this can't tell you

The exact token count of a tool call depends on the model, the client, how many servers are connected and how long the conversation is, and the clients don't show it per call. Subscription allowances are also expressed differently across Claude and ChatGPT plans and change over time. The honest guide is your plan's usage page and the service's cost display, not a fixed rule of thumb.

Try it

Connect Vi3W to Claude or ChatGPT and watch the cost before each job, or open Vi3W in the browser to see the same prices without an assistant in the loop.

Frequently asked questions

does mcp use tokens

Yes. Tool names and descriptions, the arguments the model writes and every result it reads back all go through the model's context, so they count as tokens or plan usage like any other text.

does mcp consume more tokens

It can. Each connected server adds its tool descriptions to the context, and long tool results (big scene dumps, screenshots) are read in full. Connecting fewer servers and asking for short results keeps usage down.

how to reduce mcp token

Turn off connectors you aren't using, ask for summaries instead of full dumps, avoid screenshot tools unless you need to see the scene, and start a new chat when a long session has drifted.

Last reviewed