Ardaro: a tested native FastMCP client example for receipt/invoice review

[ashevilleislove]

I'm sharing this on behalf of Ardaro, which operates the linked service. This post was prepared with AI assistance and checked against a live client run.

If you are testing a remote MCP client before connecting it to a model, Ardaro exposes two fixed synthetic receipt/invoice fixtures. They let you inspect structured document results and the explicit human-review boundary without providing a document, API key, or wallet.

This is a native fastmcp.Client example, not a claim that our server is built with FastMCP. The useful distinction is between successful discovery, successful fixture execution, and accepting the result for an accounting action: those are separate steps.

Tested September 13, 2026 with Python 3.11.0, FastMCP 4.0.3, MCP 2.2.0 and httpx2 2.12.0. In an isolated Python environment, install fastmcp==4.0.3, mcp==2.2.0 and httpx2==2.12.0, then run:

import asyncio
import hashlib
import json
from fastmcp import Client

async def main():
    async with Client("https://agents.getardaro.com/mcp",
                      mode="legacy", timeout=30) as client:
        names = {tool.name for tool in await client.list_tools()}
        examples = ("get_receipt_example", "get_invoice_matching_example")
        if not set(examples).issubset(names):
            raise RuntimeError("Expected free fixtures are missing")
        print("catalog:", len(names), "protocol:", client.protocol_version)
        for name in examples:
            result = await client.call_tool(name, {})
            fixture = result.structured_content
            if result.is_error or not fixture or fixture.get("synthetic_only") is not True:
                raise RuntimeError("Expected a successful synthetic fixture")
            response = fixture.get("expected_response", fixture.get("response"))
            if response["status"] != "review_required":
                raise RuntimeError("Expected the human-review boundary")
            canonical = json.dumps(response["result"], sort_keys=True,
                                   separators=(",", ":"), ensure_ascii=False)
            identity = "res_" + hashlib.sha256(canonical.encode("utf-8")).hexdigest()
            if response["identities"]["analysis_result_identity"] != identity:
                raise RuntimeError("Result identity mismatch")
            print(name, response["status"], identity)

asyncio.run(asyncio.wait_for(main(), timeout=90))

The recorded run discovered 18 tools from ardaro-agent-services 1.3.0, negotiated protocol 2025-11-25, and invoked exactly these two examples with {}. Both returned review_required and matching canonical result identities. The receipt example has one line item and a total of 107.00; the invoice example has one matched line and three uninvoiced PO units. Uninvoiced quantity does not establish whether goods were delivered.

Two limits matter: FastMCP warns that some JSON-schema regex lookahead patterns are not enforced by its Pydantic validator, so this is not full client-side schema validation. The hashes compare result contents; they do not prove document authenticity or authorize posting to accounting. No model, paid tool, or accounting mutation was exercised by this example. Other tools in the discovered catalog have separate contracts and payment requirements.

Service quickstart and contracts · Runnable examples and scoped evidence

For FastMCP users building financial review steps: is preserving the raw structured_content plus a separate review status the boundary you would expose to the next node, or do you require a smaller validated application model first?