Skip to main content

MCP Setup Guide & Best Practices

This guide will help you interact with Carbon Arc data using natural language through the Claude Desktop App and the CarbonArcMCP server.

Our system translates everyday business questions into structured analytical frameworks and supports discovery, helping you explore available insights and entities, even if you’re not sure what to ask yet.

Interested in Access?​

Contact us at support@carbonarc.co to request access.

Setup Guide​

Video Walkthrough​

Requirements​

  • Claude Pro subscription
  • Claude Desktop App installed
  • A Carbon Arc platform account

Setup Instructions​

  1. Open the Claude Desktop App

  2. Navigate to:
    Settings → Connectors → Add Custom Connector Settings

  3. Fill out the form:

    • Name: CarbonArcMCP
    • Remote MCP Server URL: https://mcp.carbonarc.co

  4. Click Connect Connect MCP

  5. A browser will open to authenticate. Log in with your Carbon Arc credentials

  6. Once connected, the tool will appear under Connectors

  7. Open a new chat and start asking questions like:
    "What was Walmart’s card spend in Q1 2024?"

MCP Prompt

Best Practices​

1.Ask Direct Business Questions (When You Know What You Want)​

Use this when you know what you're looking for.

Examples:

  • “What was Walmart’s card spend in 2024?”
  • “Compare Target and Costco in terms of app downloads.”

Responses include a markdown-formatted table with the resulting insights, along with metadata on selected filters and available options. This allows Claude to directly answer your question and build on the output in follow-up queries.

WMT Example

2. Explore What Type of Data Is Available (When You’re Not Sure Yet)​

If you’re exploring or not sure what data is available, try asking:

  • “What data do you have for Walmart?”
  • “What insights are available for Florida?”
  • “Which entities are included in the Card Spend insight?”

These types of questions help Claude discover the available entities and insights to guide your next step.

Data Walmart

3. Get to Know the CarbonArcMCP Server​

As you get familiar with the Carbon Arc MCP Server and tools available, try asking Claude:

  • “What tools are available within CarbonArcMCP?”
  • “Give me an example of how the text_to_insight tool works.”
  • “Use the text_to_insight tool to explore Walmart’s quarterly card spend for 2024.”

These types of questions help drive further orchestration and querying of the MCP server to generate meaningful responses.

Get To Know

4. Other Helpful Tips​

  • If you want to control the query precisely (e.g., select entity, insight, filter, aggregation), phrase your request with specifics:
    • “Give me total card spend for Costco in CA from Jan to Dec 2024, by month.”
  • All responses include structured data and often metadata. Claude will use this to answer follow-ups.
  • If you ever get a vague or empty answer, assume it might have skipped the tools. Just re-ask with “use mcp tools” or “use tool_name tool” at the end of your question to help guide Claude.
  • You can allow Claude to call the tools Allow unsupervised for a more uninterrupted conversation or Always ask for permission.

Unsupervised

Claude Allow

FAQs​

Freequently asked questions about Carbon Arc's MCP Server.

How does natural language querying work with Carbon Arc data?​

Carbon Arc provides a framework—comprising of elements such as entity, insight, time, geography, and optional filters—to help answer business questions submitted by users. Natural language inputs are transformed into structured queries designed to generate meaningful and actionable insights.​

Can I use ChatGPT or my own in-house LLM?

During the initial beta program, Carbon Arc is integrated exclusively with Claude Pro. If you have a preferred agent that you would like us to consider for future due diligence and technical integration, please contact us—we welcome your input as we expand support.​

How does Compliance and data asset-level permission work?

At present, Carbon Arc is integrated with Claude, which according to its stated policy here, does not use inputs or outputs to train its models. As a result, the initial rollout is limited to Claude Pro while additional agents undergo technical and privacy due diligence.

Will Claude cite to the underlying Carbon Arc data used to create output?

Claude can be prompted to cite the Insight ID and the Topic (e.g., Insight ID 775 – “App Downloads”, CA0054 – Core Panel) but it may or may not choose to cite the underlying data on its own. However, it will be clear if the data sourced is from Carbon Arc as the MCP tool calls are returned.

Are there any differences between the data I as a client currently receive and the data accessible via the MCP server?

The underlying data source for the MCP Server is the same as that of the Platform. Bulk / row level data is different and not surfaced via the MCP server. So generally, the same data accessible today via the Platform is accessible via the MCP Server.

How does the MCP Server handle compliance or restricted alternative data?

Alternative data refers to a different source that is not restricted. For example if the prompt is “Show me Walmart credit card spend” and card spend from Data Provider A is restricted, it will surface card spend from Data Provider B if unrestricted and available.

Is the Claude Desktop app required to use the MCP Server?

No, you can use https://claude.ai/ as well

Can Carbon Arc see the prompts clients put in to Claude Pro?

No, we can see only what Claude Pro chooses to pass to each Carbon Arc tool as per the Model Context Protocol. Claude Pro may adjust the prompt, simplify the prompt, or change the prompt entirely but Claude Pro will only share a synthesis of the prompt required to generate the response. Clients can see exactly what was made visible to Carbon Arc by expanding the tool box displayed by Claude Pro.

Is the synthesized prompt that Claude Pro shares with Carbon Arc ever shared with third parties or surfaced in a manner that other customers might be able to see?

No, we do not surface the synthesized prompt that Claude Pro chooses to share with Carbon Arc

What part of the response generated by Claude Pro can Carbon Arc see?

Carbon Arc sees no part of the generated response by Claude Pro (i.e., in the chat session). Carbon Arc can only see what the Carbon Arc MCP server surfaces back to Claude Pro.

Can Carbon Arc link back a synthesized prompt to a specific user or organization?

Yes, Carbon Arc can see the logged information by Carbon Arc user-credentials (name and email address used for Carbon Arc Platform)

What information is Carbon Arc logging for usage?

Carbon Arc logs the following information:

  1. What tool was called and when
  2. Inputs passed from Claude Pro to our tools (tool-specific parameters provided by the user’s prompts)
  3. Responses generated by Carbon Arc’s MCP Server (the outputs returned back to Claude Pro)
  4. Carbon Arc User information: name and email address associated with their Carbon Arc-authenticated API token. We do not log Claude Pro credentials.
  5. Token consumption: the number of Carbon Arc and LLM tokens used
Do outputs get reused for training any AI models?

No

Can Carbon Arc see or retain conversation history across chats?

No

Is client data shared with LLM providers (e.g., OpenAI, Anthropic)?

No