MCP Integration

UserTold captures consented in-product interviews: voice and transcripts, in-page behavior and page context, plus screen sharing when a participant approves it in a supported desktop browser.

Interviews produce source-linked Evidence from what participants said, did, and saw; related Evidence becomes draft Findings for review. Generated transcripts and analysis do not replace the captured source.

UserTold does not recruit participants or replace consent. Mobile and unsupported browsers continue without screen sharing. Denied permissions, interruptions, and failures can leave capture gaps.

Fast path: let your agent work with existing research

Using ChatGPT or Codex? Install the UserTold plugin.

For other MCP clients, connect directly:

Help me connect UserTold to this agent with https://mcp.usertold.ai/mcp and complete browser authorization.

Discover the organization and Project bootstrap resources. Ask me to select or create a Project, then proceed. Propose changes before acting.

Server:

https://mcp.usertold.ai/mcp

Clients use OAuth 2.1 browser authorization with PKCE; no API token is needed.

Agents can inspect grounded Study results and Evidence, download authorized Interview artifacts through short-lived links, maintain Findings, and explicitly deliver ready Findings. Bytes are never inlined; private administration stays in the dashboard.

Operate safely

Propose Study changes before acting. Activating a Study starts participant capture; closing stops new interviews without interrupting one already in progress.
Start completed-research review with studies.get_results; follow its references with evidence.get and interviews.get_context, or use interviews.get_artifacts when exact source bytes are required.
Before marking Finding ready, read its evidence with findings.get_evidence and the interview behind it with interviews.get_context.
Read interviews.processing_status before calling interviews.retry_processing, and never retry on a loop.
Delivery sends ready Findings to an explicit GitHub or Linear destination; it does not decide readiness.
If a UserTold MCP tool is broken, confusing, or behaves differently from what you expected, call feedback.submit with a concise safe reproduction. Do not include credentials, participant identity, personal data, or raw private research content.

If setup fails

Reconnect authorization. For capture failures, run projects.verify_widget_installation. It checks the page, not the running app — to see why the widget never started for a visitor, listen for the usertold:integration-error event described in Widget Integration.

Self-discovery and next steps

Use authenticated tools/list and resources/list; their responses own the current inventory.