MCP tools and Skills
The MCP server gives an agent the tools. A User Intuition Skill gives it the workflow: when to ask for Panel versus BYOP recruitment, how to conduct the Customize Plan conversation, when to dry-run cost, and when confirmation is required.Connect the MCP server
Install the ChatGPT plugin, or configure Claude, Codex, Cursor, or VS Code, and verify the connection.
Install a research Skill
Teach your agent a repeatable study-creation or analysis workflow.
One surface on every transport
The Claude directory connector and the ChatGPT plugin use the full/mcp endpoint with OAuth sign-in.
- Streamable HTTP accepts OAuth for interactive clients and organization API keys for headless clients. Connect to
https://mcp.userintuition.ai/mcp. - stdio remains available for Claude Desktop, Cursor, Claude Code, VS Code, and other local clients. It authenticates with a
ui_sk_API key.
Smaller profiles
For the common study → interview → report workflow, connect tohttps://mcp.userintuition.ai/core/mcp. It omits webhook, feasibility, and external-panel administration; organization selection is included. For existing research without write actions, connect to https://mcp.userintuition.ai/read-only/mcp. The https://mcp.userintuition.ai/compact/mcp profile keeps the research workflow and groups less common webhook, feasibility, and external-panel operations into grouped tools. Each URL has its own OAuth protected-resource metadata, so authorize the specific URL your client uses. The full /mcp URL remains available for every tool.
Local stdio clients can pass --profile core, --profile read-only, or --profile compact after the package name. The default is the full profile. Resource and prompt support remains available on these profiles.
Measured from the complete tools/list response for MCP server v0.17.1:
The token estimate divides bytes by four; actual model context usage varies by host and tokenizer. CI checks each profile against its payload budget.
Tool groups
The tool reference documents every registered tool. The study-creation playbook documents the behavioral rules agents should follow.
The server also exposes attachable study and report resources for clients that show MCP resources in the conversation. These resources read the same authenticated public API as the research tools.
Safety model
The server advertises MCP safety annotations for every tool. Agents should also follow these product rules:- Use
customize_studyfor the plan, audience, screeners, and concept material; do not reconstruct those fields in the MCP host. - Return the complete persisted study plan and obtain approval of its current version before creating BYOP participants, sharing an external-panel entry link, or launching paid Panel recruitment.
- Use
estimate_panel, show the complete estimate, and carry itsestimate_idinto an approved paid panel launch. - Require explicit user confirmation before rewards, webhooks, stopping fieldwork, or deleting studies, participants, or interviews.

