Compare AI model costs in Cursor and Claude with the MCC MCP server

Model Cost Comparison’s MCP server lets a compatible AI assistant retrieve model pricing and estimate token costs during a conversation. You can ask for a comparison while working on an application, then inspect the rates and assumptions behind the result.

We built modelcostcomparison.com at Lazige. This integration uses that site’s pricing catalogue and calculation engine. The assistant decides which tools to request. The tools return structured pricing and calculation results. That gives the conversation a source to consult when answering cost questions.

This walkthrough uses an illustrative support-assistant workload. It is a setup guide, not a report of customer savings or measured production performance.

Connect your AI client

The remote server address is:

https://modelcostcomparison.com/api/v1/mcp

The transport is Streamable HTTP. MCC does not require a provider API key to use these pricing tools. Your AI client may have its own account requirements or usage charges.

Cursor

Open the MCC developer page and use Install in Cursor, or merge the following entry into your existing MCP configuration:

{
  "mcpServers": {
    "modelcostcomparison": {
      "url": "https://modelcostcomparison.com/api/v1/mcp"
    }
  }
}

Keep any other servers already in your configuration. Enable the connection and check that the client discovers the available tools before you start the comparison.

Claude

In a Claude environment that supports custom remote connectors, add a connector named ModelCostComparison and use the server address above. Enable its tools in the conversation. Your organization may require an administrator to add or approve the connector.

Follow the maintained setup instructions if your client’s settings differ. A connection makes the tools available. It does not mean the assistant will call them in every conversation.

What the six tools do

Discovery comes first because a familiar model name can correspond to several offers. Use the identifiers returned by MCC rather than asking the assistant to invent them.

Work through a realistic comparison

Suppose you are planning a support assistant with 50,000 requests a month. For this example, assume 800 input tokens and 200 output tokens per request, with no caching. That becomes 40 million input tokens and 10 million output tokens monthly.

Use ModelCostComparison to find eligible standard-tier offers for 40 million input tokens and 10 million output tokens per month, with no cached input. Assume no request exceeds 800 input tokens. Show five low-cost options, their exact offer IDs, price-review information and any exclusions. Do not infer answer quality from price.

The client can use find_cheapest_models to search eligible offers and get_model_pricing to examine an individual result. Ask it to identify which tools it actually used, and keep the assumptions alongside the numbers.

Then narrow the comparison:

Compare three of those offers using the same workload. Separate input and output costs, include source information, and explain any offer that could not be calculated.

The output should support a decision you can inspect: which offer was used, which usage values were supplied, what the calculation includes, and where the evidence is limited. Verify that the assistant actually called the tools before you treat its answer as an MCC result.

Add caching carefully

After a first comparison, you may want to test a caching scenario:

Repeat the comparison with 10 million of the 40 million input tokens treated as cached input. Keep total input and output unchanged. Explain whether each offer supports that assumption.

Cached input is part of total input, not an extra token quantity to add on top. In this example, the split becomes 30 million uncached input tokens and 10 million cached input tokens.

This remains a scenario until your application demonstrates that cache usage. The MCP subtotals exclude cache-write charges and other separately billed services, so a discounted cached-read rate alone does not describe the full economics of caching.

Read the boundaries of the result

MCC’s MCP calculations cover text-token subtotals. Search, tools, storage, cache writes, and taxes are outside that subtotal. An estimate is not a complete invoice forecast. How those boundaries are applied is documented in the ModelCostComparison methodology.

The catalogue is also a subset of available offers. The server can reject calculations for expired or unreviewed entries rather than presenting an unsupported price. Cheapest-offer results reflect the eligible catalogue and the filters you supplied, not every model on the market.

A per-request input-capacity filter uses reviewed pricing input limits. It does not establish the model’s full context-window capability. Entries with unknown limits are excluded when that filter is used.

Cost ranking does not measure answer quality. Use the website’s benchmark information and your own task evaluation to decide whether a low-cost candidate meets the application’s requirements. The calculator, the benchmarks, and the models advisor are covered in Model Cost Comparison: choose AI models around your workload.

If the connection works but no comparison appears

Check that the connector is enabled for the current conversation, then explicitly ask the assistant to use ModelCostComparison. Seeing tools listed confirms discovery. It is not evidence of a completed calculation.

For invalid-parameter errors, ask the assistant to read the tool schema and use an exact offer identifier from list_models. For transport or protocol errors, check the client’s MCP support and the current setup instructions. Do not put database credentials or administrative tokens into the public connector configuration.

Try one decision before automating more

Begin with a single workload and a small shortlist. Save the assumptions and review information, test the candidates on real examples, and repeat the comparison when usage or available offers change.

Connection settings, kept apart from the setup guide: modelcostcomparison-mcp.