You are helping the user put an AI assistant on top of Helio via the Model Context Protocol — Claude Desktop, Cursor, Claude Code, custom agents — for both reading research data and building tests.

Core idea

Helio MCP exposes the platform to any MCP-aware client as 20 tools, 2 prompts, and 1 resource, with auth at the protocol layer. The tools group into:

Discovery (5) — list_projects, get_project, get_project_tests, list_tests, get_test
Building a draft (5) — create_test (10 question types, 11 UX metrics, ux_metric_context, ux_metric_assets), add_question, update_question, remove_question, update_test (metadata + metric add/remove + reorder, bundled)
Launching (3) — validate_test, send_test, delete_test
Reading (4) — get_test_report (the workhorse, with include sections and demographic filters), get_filtered_responses, get_responses, compare_segments
Assessments (3) — list_assessments, get_assessment, create_assessment

Plus the analyze_test and compare_audiences prompts and the helio://guide/analysis-workflow resource.

Build the measurement spine from ux_metrics, not hand-written questions. A tagged metric auto-builds its sections with validated, non-leading wording and returns a 0–100 score with a threshold label, comparable across waves and rolled into the Overall Score; a hand-written question that resembles a metric returns only an answer distribution and feeds none of that. Pass questions for what no metric covers (click tests, MaxDiff, risk-specific probes). MCP has an edge here the CLI lacks — ux_metric_assets attaches an asset per metric at create time. See helio-ux-metrics for the argument and helio-patterns for which metric set fits the stage.

Two transports, same contract: stdio (npx @zurb/helio-mcp — Claude Desktop, Cursor, Claude Code) and hosted HTTP (https://mcp.helio.app/api/mcp with a Bearer token — Claude.ai web, custom agents).

Ceilings: no asset upload or listing on the MCP surface (the CLI has assets upload/list/get since v0.1.1 — route image uploads there), no click/tree/prototype creation, no branching, no audience creation, plus two MCP-specific gaps: no dry-run (validate_test is the pre-spend check) and no participant-eye walkthrough (use the preview URL create_test returns).