docreadi-mcp 0.1.0

MCP server for DocReadi — document data extraction for finance, in your AI client.

docreadi-mcp

A local MCP server for DocReadi — document data extraction for finance. It runs on your machine, reads documents off your disk, and calls the hosted DocReadi API with your API key, so an AI client (Claude Desktop, Claude Code, Cursor, …) can extract documents and query your DocReadi corpus without leaving the chat.

Status: v1, in progress (see ../../MCP_SERVER_PLAN.md). Tools land phase by phase; this is the skeleton + first read tool.

How it works

Your MCP client spawns docreadi-mcp as a local subprocess and talks to it over stdio. The server is a thin courier — all extraction/storage happens on the hosted DocReadi API; the server just translates tool calls into HTTPS requests and (for ingestion) reads local files. Your API key lives only in your client config, never in this repo.

Install / configure

Add the server to your MCP client config with your DocReadi API key (get one at docreadi.com → Settings → API keys):

{
  "mcpServers": {
    "docreadi": {
      "command": "uvx",
      "args": ["docreadi-mcp"],
      "env": {
        "DOCREADI_API_KEY": "dr_live_…"
        // "DOCREADI_BASE_URL": "https://api.docreadi.com"  // override if self-hosting
      }
    }
  }
}

Env var	Required	Default
DOCREADI_API_KEY	yes	—
DOCREADI_BASE_URL	no	https://api.docreadi.com

Tools

Tool	What it does
check_connection	First-run self-test — is DocReadi reachable and is your API key valid? Never errors; returns a diagnostic.
extract_document	Extract structured data from a local PDF/JPG/PNG — uploads, waits for extraction, returns the fields + line items + totals + confidence.
get_document	Fetch a document's status + extracted data by id.
extract_adhoc	One-off, non-destructive extraction of caller-defined fields from a document already in DocReadi (e.g. "who signed these delivery notes?").
classify_document	(Re)classify a document's type; returns the type + reasoning.
search_documents	Search/list the workspace's documents (by vendor, number, type, status, or date range) — a summary row per match.
list_counterparties	List vendors/customers (VAT, aliases, doc counts), paginated.
list_reports	List the workspace's saved reports.
run_report	Run a saved report and return its rows — the way to query/search the corpus.
export_report_csv	Run a saved report and write the result as a CSV to a local path.

(search_documents is the raw query surface; for richer column selection and saved queries, author a report in the UI and use run_report / export_report_csv.)

Development

This package lives in the DocReadi monorepo so it's reviewed and CI'd alongside the API it wraps. Its layering keeps the dependency-free parts testable in the main CI without the mcp SDK:

• docreadi_mcp/config.py — env config.
• docreadi_mcp/catalog.py — pure-data tool→endpoint registry (the maintenance anchor: tests/test_mcp_catalog_contract.py binds it to agent_guide.GUIDE).
• docreadi_mcp/client.py — httpx client (tested with MockTransport).
• docreadi_mcp/server.py — thin FastMCP glue (lazy-imports the mcp SDK).

# from the repo root
pip install -e services/mcp          # installs mcp + httpx
DOCREADI_API_KEY=dr_… python -m docreadi_mcp   # run over stdio

Maintenance contract: when a DocReadi API endpoint a tool wraps changes, update docreadi_mcp/catalog.py (and the tool) — the contract test fails CI otherwise. See MCP_SERVER_PLAN.md §5.