Backroom MCP Server

mcp-name: io.github.pardhuparimi/backroom

An MCP (Model Context Protocol) server that wraps the Backroom invoice-extraction API, so any MCP client can submit invoice PDFs and read back structured, verified extractions as tools.

Backroom never guesses: every field in a proposal carries a value, a verification status (verified / extracted / unclear), a confidence, and provenance. A field the engine could not verify comes back unclear with a null value instead of a fabricated one.

Install & run

With uv:

uvx backroom-mcp

Or with pip:

pip install backroom-mcp
backroom-mcp

The server speaks MCP over stdio — your client launches it as a subprocess. Example client configuration:

{
  "mcpServers": {
    "backroom": {
      "command": "uvx",
      "args": ["backroom-mcp"],
      "env": {
        "BACKROOM_API_KEY": "bk_test_…"
      }
    }
  }
}

Configuration

Tools

⚠ Proposals are fetch-exactly-once

The API runs zero-retention by default: the first successful get_proposal fetch purges the result, and a second fetch of the same id returns 404. Whatever client calls it must keep the returned proposal — it cannot be fetched again.

Errors

The API reports errors as RFC 9457 problem documents (title, code, detail, retryable, request_id). This server surfaces them verbatim as the tool error — including reason codes such as NO_EXTRACTABLE_TEXT when a document is a scan with no readable text, which the API refuses rather than guesses at.

Zero retention

This wrapper stores nothing. PDFs are read from disk at call time and sent to the API over HTTPS; no document, result, or key is written to disk or logged by the server. On the API side, zero-retention keys purge results on first fetch or after 24 hours.

Development

python -m venv .venv && .venv/bin/pip install -e ".[dev]"
.venv/bin/python -m pytest

Tests mock the HTTP layer — no network and no API key needed.