The MCP surface grows from 8 to 15 verbs and becomes self-describing, without moving the fence: every verb is still navigation or visualization, none returns a finding, none runs code, nothing leaves the machine.
- viewer_state: what is on screen, machine-readable, with every slice in the pane count and its direction.
- pick_scan: the agent asks, you choose in the native dialog; the agent gets the label, never the path.
- Saved views over MCP (list, restore, save); agent-saved views carry an "agent" badge so a bookmark is never mistaken for yours.
- set_3d_style, navigate_arch (pano), and slice navigation in grid and pano.
- Tool annotations and output schemas on every verb; snapshots carry a state caption.
- Read-only resources and four reading choreographies as prompts (mpr-survey, pano-survey, tmj-read, replay-views), distilled from the reading guides.
- One-click install for Claude Desktop: the attached cbctscope-1.6.0.mcpb (installs the agent half; the viewer still runs from this repository).
- Agent Skill in skills/cbctscope-reading/ for any host that reads SKILL.md.
Docs: docs/mcp.md and manual chapter 9.