MCP server

Your users’ own words, in the agent you already use.

Every conversation the widget records is readable over MCP. Ask in your own words, the way you would ask the person who sat in on all the calls.

Address

Transport

http

Streamable HTTP. No local process, nothing to install, nothing to keep running.

Authentication

OAuth 2.1

OAuth 2.1 with dynamic client registration. Your client opens a browser, you sign in with Google, and it holds a token scoped to your account. There is no API key to create, paste or rotate.

Scope

per account

One account. A token reaches the projects, placements and conversations of the account that signed in, and nothing else.

Connect it

Four clients, each with the steps that work for it and the one value those steps need.

What to ask it

The 8 tools

What a connected client can do, and whether it only reads. This table is generated from the server’s own catalog, so it cannot claim less than the server grants.

Feedback

  • feedback_listRead

    See recent sessions — where each one happened, how long it ran and how it ended. Not what was said.

  • feedback_getRead

    Read a full transcript, turn by turn, in the words the person actually used.

Placements

  • placements_listRead

    See where the agent is being offered across your products, and what it asks in each place.

  • placements_createWrite

    Name a new place to collect feedback, before the first session arrives there.

  • placements_updateWrite

    Rename a placement, or rewrite the brief and opening line every future visitor there will hear.

Projects

  • projects_listRead

    See your projects, the client id each one embeds, and which sites are allowed to use it.

  • projects_createWrite

    Add a project and its client id. It allows no sites at first, so it cannot collect anything yet.

  • projects_updateWrite

    Rename a project, or change which sites are allowed to start a session on it.

What it does with your data

A person consents before the microphone opens, and their voice never reaches our servers — it streams from their browser to the model, and audio is not stored anywhere, at all. What we keep is the transcript. A connected client reads the transcripts belonging to your account and nothing belonging to anyone else. The full terms are in the privacy policy and terms.

Questions

The projects, placements and conversation transcripts belonging to the account that signed in. A token is scoped to one account, so a client can never reach another customer’s data. Half the tools only read; the four that write are named on this page.

Yes. The card says it will record and transcribe before anything starts, and the microphone does not open until they press it. That consent is the legal basis for the recording, which is why the card will not start without it.

No. A visitor’s voice streams from their browser straight to the model that runs the conversation — it does not pass through our servers, and we do not store audio anywhere. The transcript is what gets kept and what you read over MCP.

No. We do not sell your data and we do not use transcripts to train machine learning models. The model running a conversation is given the brief for that placement and hears that conversation, and is given no other customer’s data.

Yes. Each client registers itself and holds its own token, so Claude on your desktop, Claude Code in your terminal and Cursor can all be connected to the same account at the same time. Signing in again in one does not sign the others out.

Not for the MCP server — it takes an address and a sign-in. Collecting the feedback in the first place needs the widget on a page of your product, which is one script tag or one React component. The docs carry both.