@triscope/mcp

The triscope MCP server (stdio JSON-RPC) — gives an AI agent structured, machine-readable control of a running @triscope/core lab over a headless Chromium/WebGPU pool.

What it exposes

Tools for the full converge-a-3D-scene loop, including:

• capture_views / capture_motion — render the lab's cameras to PNGs (the only reliable WebGPU readback path) with GPU + flat/black-frame probes.
• inspect_scene / read_uniform / set_uniform — scene-graph introspection and live uniform/material/light read-write.
• read_telemetry / set_knob / list_elements — observe and tune, with per-call devUrl/labUrl so one server can target any running lab.
• set_reference / diff_reference / check_targets — per-tile SSIM
  • diff heatmaps and target gating.
• add_element / remove_element / set_scene_param — drive a runSceneLab scene live.
• auto_tune / multi_tune, snapshot / restore, import_element, scaffold_from_gltf.

Run

// in your MCP client config
{ "command": "npx", "args": ["-y", "@triscope/mcp"] }

Part of the triscope monorepo. MIT.