Add a minimal Python MCP client example

[ShlokShar]

Add a small runnable Python example that connects to the hosted MCP server, calls search_decisions, and prints the returned German citation strings.

Fixes #62

[jonashertner]

Thank you for this — first contribution to the new issue set, and the structure is right: it's small, runnable, has no dependency on a local corpus, and the README link lands in the correct section.

I ran it against production before reviewing, and it exits 0 but prints nothing. The cause is a genuine wrinkle in our API that the issue text didn't warn you about, so this is our documentation gap rather than a mistake on your part.

search_decisions returns formatted Markdown text, not JSON, and no structuredContent:

content types: ['TextContent']
structuredContent: none
text begins: 'Found 1206 decisions (showing 1–3):\n\n**1.** [BGE 125 III 70](https://...) (1998-10-13) [bge] [de]'

So json.loads(text) never runs (the text doesn't start with { or [), _extract_results receives a plain string and returns [], and the loop prints nothing.

Two ways to fix it, and I'd prefer the second:

Option A — print the text. Drop the JSON handling and print the tool's own formatted output. Honest and three lines shorter, but it doesn't demonstrate reading a specific field, which was the point of the example.

Option B — use cite for the citation strings. Parse the decision links out of the search text, then call cite for each one. cite does return JSON, with a citation_string_de field. That also demonstrates the rule that matters most here: citation strings are copied from a stored field, never assembled. This is what I'd merge.

Either way, two smaller points:

1. Endpoint. sse_client needs the /sse path — https://mcp.opencaselaw.ch/sse, not the bare host. It happens to work today, but the documented endpoints are /mcp (streamable-http, preferred) and /sse.
2. Silent success. Consider exiting non-zero, or printing a message, when zero citations are found. A script that prints nothing and returns 0 is hard to distinguish from a broken connection — which is exactly what happened here.

I've also opened this as a documentation fix on our side: the text-versus-JSON split across tools isn't written down anywhere, and you're the second consumer to hit it.

Happy to take this in either shape — say which you'd prefer and I'll review promptly.

[jonashertner]

Thank you for this — and for picking up a good first issue. The structure is right and the transport works. One thing blocks it, which I verified by running your file as-is:

$ python examples/python/minimal_mcp_client.py "missbräuchliche Kündigung"
$ echo $?
0

Exit 0, no output. The cause is that search_decisions returns Markdown, not JSON, so _extract_results finds no results/items/decisions/data key and the loop has nothing to print. There is a second, smaller issue behind it: _extract_content_text handles str and dict, but the SDK yields TextContent objects, so text is empty and the code falls through to model_dump(), which JSON-dumps the envelope rather than the payload.

Two ways to finish it, either is fine:

1. Keep search_decisions and print what it returns — add getattr(part, "text", "") to _extract_content_text and print the Markdown directly. Simplest, and honest about the tool’s shape.
2. Switch the demo to a JSON tool so the "print the citation strings" promise holds literally. cite returns a JSON object with citation_string_de/_fr/_it, e.g. cite(reference="bger_6B_267_2012").

I would take (2) for this example, since printing citation strings is what it advertises, and it doubles as a demonstration of the R1 contract (copy citation_string verbatim, never construct it). Which tools return text and which return JSON is genuinely under-documented — that is #73, and your PR is a good argument for prioritising it.

Happy to merge once it prints something. No rush.