Testing a server¶
Call tools in-process. await server.call_tool("notes.search", {"query": "x"}) returns the same result a client gets, including _meta.ouncemcp with tokens and error categories.
Test as a verified caller.
from ouncemcp.testing import as_caller
with as_caller(subject="alice", tenant="acme", scopes=["tickets:write"]):
result = await server.call_tool("tickets.close", {"ticket_id": "T-1"})
Run the security catalog (EVAL-2). It always probes guessed names for every tool with a policy, plus the cases you add:
from ouncemcp.eval import CrossTenant, CursorReuse, HandleReplay, SSRFProbe, run_security_suite
ACME, GLOBEX = {"subject": "a", "tenant": "acme"}, {"subject": "m", "tenant": "globex"}
report = await run_security_suite(server, [
CrossTenant("invoices.list", {}, owner=ACME, intruder=GLOBEX),
CursorReuse("invoices.list", {"limit": 1}, owner=ACME, intruder=GLOBEX),
HandleReplay("report.start", {}, "report.approve", "handle", "handle", ACME, GLOBEX, single_use=True),
SSRFProbe("web.fetch", "url"),
])
assert report.passed, report.render_text()
TokenPassthrough(tool, args, recorded=...) checks that recorded upstream requests never carry the caller's token.
Protocol conformance (EVAL-1). python -m ouncemcp.eval.conformance serves the fixture tools the official MCP conformance suite calls; .github/workflows/conformance.yml runs the suite for MCP 2026-07-28.
Budgets in CI. Run ouncemcp check server.py; it exits non-zero when a definition is over its fail threshold or a lint error is found.