The real-harness run: Claude Code drives the native app

2026-11-11, evening. The check owed since review-tour.md. Mark refreshed the CLI login; this ran as soon as it was done.

How it ran

• The app: QPOINT_DATA=synthetic, qnative b4cbc23, the default world (nina-mbp's quiet laptop). It logged its endpoint and this launch's token:

  guide: MCP on http://127.0.0.1:7891/mcp (token …)

• The harness: Claude Code 2.1.322, headless.
  • claude -p "<the prompt>" --mcp-config native-guide.mcp.json --strict-mcp-config --allowedTools mcp__qpoint-guide-native --model sonnet.
  • The config names one HTTP server with the Authorization: Bearer <token> header.
  • --mcp-config was used rather than claude mcp add, so nothing was left in the user's Claude config. To keep the server, paste the line Help › Getting started shows: claude mcp add --transport http qpoint-guide-native http://127.0.0.1:7891/mcp --header "Authorization: Bearer <token>".
• The prompt, as a person would ask it: "I just opened Qpoint on this laptop. Give me the quick tour of what's on this machine, then show me which model costs the most and where I'd see that."

What it did

The server connected on the first try, with the token. Ten turns, 18 seconds, $0.13 (Sonnet 5.5), no tool errors:

1. tour-state, list-tours
2. find-in-app "which models cost the most"
3. start-tour this-machine, then read-step open (the live counts)
4. open-page models (the sidebar's Models ringed first, then the page)
5. point models.list with a title, read the region's text, then point again with a title and one sentence

What it said

• The tour: the opening's numbers for the last 30 days.
  • 2 providers, 3 harnesses, 4 models.
  • 5 MCP servers, 22 tools, 3 skills.
  • 1 workspace.
  • It said the tour had started and named its stops.
• The most expensive model: "You'll see this under Inventory › Models" (opened and ringed). Then a table read from the rows' own text:

  Model	Tokens	Spend
  claude-sonnet-4-6	5.12M	$101.98
  claude-opus-4-8	504k	$48.90
  claude-haiku-4-5	1.90M	$11.42
  gpt-5.5	661k	$3.22

• It added that Sonnet has the most spend and Opus costs the most per token (~$97 per million tokens).
• It pointed at the band's top three by cost and said a row opens the model.
• Nothing was pressed.

The window afterwards

tour-state, read over the same endpoint:

• the page is /inventory/models;
• the pointer is on models.list: "Spend per model · Each row shows tokens and spend; claude-sonnet-4-6 is the biggest total, claude-opus-4-8 the priciest per token.";
• the last move is { by: agent, driver: claude-code, walk: true };
• the tour is still at its opening step, and the panel says the step is on another page.

The W7 criterion the lab set for its own guide ("Claude Code walks the tour from one prompt") holds for the native app.