Receipt Audit

Receipt fraud detection MCP server

Receipt Audit's MCP server lets an AI agent check a receipt or invoice for edits, AI generation and reuse, read its fields, and get its own API key. The endpoint is https://receiptaudit.com/mcp (Streamable HTTP). There is no sign-in: an agent creates a key with one tool call.

Install

Claude Code

claude mcp add --transport http receipt-audit https://receiptaudit.com/mcp

To use the same account in every session, add your key as a header:

claude mcp add --transport http receipt-audit https://receiptaudit.com/mcp \
  --header "Authorization: Bearer raud_…"

Claude Code also reads the .mcp.json shown for VS Code below. Claude Code MCP docs

Claude and Claude Desktop (custom connector)

  1. Open Customize, then Connectors. Click "+ Add", then "Add custom connector".
  2. Name it Receipt Audit, enter https://receiptaudit.com/mcp and click Continue.
  3. Under Authentication, choose "No sign in".
  4. Optional: under Request headers, add authorization with the value Bearer raud_…. Anthropic says request headers are in beta for a limited set of organizations.
  5. Click Add. Turn the connector on in a chat from the "+" button, then Connectors.

Without a header, Claude calls create_api_key and passes the key as the api_key argument. Keep that key (in a project's instructions, say) so later chats use the same account. On Team and Enterprise plans an owner adds the connector first, in Organization settings. Free plans allow one custom connector. Sources: custom connectors, request headers.

Cursor

Add this to ~/.cursor/mcp.json for every project, or to .cursor/mcp.json for one. Leave out headers to start without a key. Cursor MCP docs

{
  "mcpServers": {
    "receipt-audit": {
      "url": "https://receiptaudit.com/mcp",
      "headers": {
        "Authorization": "Bearer raud_…"
      }
    }
  }
}

VS Code

Run MCP: Add Server, choose HTTP and enter https://receiptaudit.com/mcp. Or add this to .mcp.json at the workspace root, or to ~/.copilot/mcp-config.json for every workspace:

{
  "mcpServers": {
    "receipt-audit": {
      "type": "http",
      "url": "https://receiptaudit.com/mcp"
    }
  }
}

To use one key everywhere, add "headers": { "Authorization": "Bearer raud_…" }, and don't commit a file that holds a key. VS Code MCP docs

Tools

Listed from the server's own definitions, which also serve the REST API. A tool that needs a key reads it from the Authorization header, or from an api_key argument for clients that can't send headers.

check_receipt: Check a receipt

Extracts the fields of a receipt or invoice (image or PDF, by https URL or base64), checks its arithmetic, tax rate and dates, scores signs that it was edited or AI-generated (with reasons), and reports whether the same receipt was submitted before. Costs 1 credit, refunded if the file can't be read.

REST: POST /api/v1/receipts/check · needs a key

  • image_url: Public https URL of the receipt image or PDF.
  • image_base64: The file's bytes, base64-encoded (a data: URL prefix is fine). Max 8 MB.
  • mime_type: Only needed with image_base64 when the type can't be sniffed. One of: image/jpeg, image/png, image/webp, image/heic, image/heif, application/pdf.
  • api_key: For MCP clients that can't send an Authorization header. Prefer the header.

create_api_key: Create an API key

Returns a new API key with 50 free checks, and a claim_url where a person can take over the account and add prepaid credits. One key serves every later call.

REST: POST /api/v1/keys · no key needed

No parameters.

get_account: Get account and credits

Returns the remaining credits, whether a person has claimed the account, and how many checks it has run.

REST: GET /api/v1/account · needs a key · read-only

  • api_key: For MCP clients that can't send an Authorization header. Prefer the header.

report_feedback: Report feedback

Report a missing capability, a wrong result or a bug in a sentence or two. A person reads every report. Free, no key needed (send yours to link the report to your account).

REST: POST /api/v1/feedback · no key needed

  • kind (required): One of: missing_capability, wrong_result, bug, other.
  • message (required): What you were trying to do, what happened, what you needed.
  • operation: The operation it's about, e.g. check_receipt.
  • api_key: For MCP clients that can't send an Authorization header. Prefer the header.

create_claim_link: Create a claim link

Returns a new claim_url (valid 30 days) where a person can claim this account and buy credits.

REST: POST /api/v1/account/link · needs a key

  • api_key: For MCP clients that can't send an Authorization header. Prefer the header.

The server also lists one resource, benchmarks://receipt-audit/latest: The latest published RA-Bench run summary, the same JSON as https://receiptaudit.com/benchmarks.json. Run by Godard Builds, which builds Receipt Audit.

No sign-up: the agent gets its own key

  1. The agent calls create_api_key. It gets an api_key with 50 free checks and a claim_url.
  2. It stores the key and sends it with every call, as a header or as the api_key argument.
  3. It gives the claim_url to its person. The link works for 30 days. They open it, enter an email and buy a credit pack (from $49 for 3,000 checks). The credits land on the same key.
  4. When credits run out, check_receipt fails with payment_required and a fresh claim_url. create_claim_link makes a new link at any time, and get_account shows the balance.

Example prompts

Prompts with an https link work best. Sending a file as base64 makes the agent write the whole file into its tool call.

Questions

Does the MCP server need OAuth?

No, it has no sign-in. Tools that act on an account take an API key, and an agent can create one itself with create_api_key: 50 free checks, no sign-up.

What does it cost?

Each new key comes with 50 free checks. After that, prepaid credits from 1.6¢ a check ($49 for 3,000 checks). Only check_receipt uses credits: one per check, refunded if the file can't be read.

Which transport does it use?

Streamable HTTP, at https://receiptaudit.com/mcp, with no sign-in.

Machine-readable descriptions