Reference MCP client for Kontor MCP: shows the protocol from the client side (introspection), gives you a scriptable one-shot audit without any LLM, and an Anthropic agent loop with every Kontor tool bridged in and each tool call traced.

kontor-agent tools                                   # tools / resources / prompts with schemas
kontor-agent audit invoice.xml [--lang en] [--known R-1,R-2] [--json]
kontor-agent chat [-m "Prüfe /abs/path/invoice.pdf"] [--model claude-sonnet-5] [--effort high]

Transport is a global option: by default the bundled @kontor-mcp/server is spawned over stdio; --url http://host:3333/mcp --token … (or KONTOR_AUTH_TOKEN) talks to a running Streamable HTTP host, e.g. the Docker image. --stdio <command…> spawns any other stdio server.

Command	Needs	Exit code
`tools`	a server	0
`audit <file>`	a server — no API key	0 accept · 1 review · 2 reject · 3 error — usable in scripts and CI
`chat`	a server + `ANTHROPIC_API_KEY`	0

chat uses the Anthropic SDK's Tool Runner (@anthropic-ai/sdk/helpers/beta/mcp bridges MCP tool definitions 1:1 — the JSON Schemas the server publishes are what Claude sees). The trace lines are the demo of the protocol:

[kontor-agent 0.9.0 · stdio → node …/server/dist/bin.js · claude-sonnet-5 · effort high]
  → audit_invoice file_path=/…/broken-zugferd.pdf lang=de
  ← audit_invoice recommendation=reject · verdict=invalid · 4 findings · 412 ms
Die Rechnung ist abzulehnen: …
  [3 412 in / 388 out · stop=end_turn]

Defaults: claude-sonnet-5 (--model claude-opus-5 recommended for hard cases — ambiguous multi-invoice questions, long reasoning over many findings), adaptive thinking, effort high, 20 tool rounds per message, streaming; the server-side refusal fallback is enabled when an Opus 5 / Fable 5 model is selected. Interactive mode keeps the full exchange (including tool calls) as history; an empty line or Ctrl-D ends the session.

Sovereignty note: the Kontor server never touches the network. chat is the optional client that talks to Anthropic — what leaves your machine is exactly the tool schemas, your messages and the tool results shown in the trace. tools and audit send nothing anywhere.

See docs/demo/ for recorded transcripts over both transports.