Referna opens its meeting recaps to AI agents through MCP

Two tools and one endpoint. That is all Referna's new MCP connector exposes. Referna, a referral network for independent professionals, announced on 14 October that its members can now give their personal AI agent read-only access to their meeting recaps. The news was picked up by AiThority from a release distributed by EIN Presswire.

It is not a large-scale launch, and that is exactly why we find it interesting: it is the kind of small, well-scoped integration with carefully designed permissions that shows where MCP is heading as it moves beyond big SaaS platforms into niche products.

What Referna has launched

Referna organises independent professionals into peer circles that serve similar clients without competing directly with each other. Those circles meet live, and the platform then generates a personalised recap for each attendee, with their own summary and coverage of the agenda items.

The change is that this material is no longer locked inside the Referna website. The member creates an integration key in their account settings and connects any MCP-compatible client, or uses the HTTP API if they prefer to build their own workflow. According to the connector listing on Glama, the server lives at `https://referna.com/mcp/member-recaps`, uses Streamable HTTP transport with the 2025-11-25 version of the specification, and offers two tools:

1. `list_meeting_recaps`, which returns the key owner's recaps sorted from newest to oldest, with pagination.
2. `get_meeting_recap`, which retrieves a specific recap by session ID.

With that, an agent can help review previous conversations, prepare for a follow-up meeting or find topics the user wanted to revisit.

The limits, in plain sight

What we like most about the announcement is what Referna leaves out. The integration only gives access to the member's own recaps, and every request checks two things: that the person attended the meeting and that they were a recipient of the recap. No recordings or full transcripts are exposed, and the key can be replaced or revoked at any time. The company also warns that the current format does not guarantee a separate list of action items, an honest detail that many announcements would have left out.

Matan Protter, co-founder, CTO and head of product, sums it up in the release published by AiThority: "We believe personal AI agents will become part of how people manage their day. An AI-native network has to give those agents access to meeting recaps through MCP, with members in control."

Why it matters

The pattern is not new in the meeting notes space: tools such as Avoma, Granola or Circleback already offer MCP servers so that Claude or other assistants can query call history. What is different here is the context. Referna is not a productivity tool but a professional network, and the information it handles (who refers whom, what each client needs) is sensitive by definition. Choosing a minimal, read-only scope is the right call for that kind of data.

It is also a good example of MCP server design. Two tools with clear names and precise descriptions are easier for a model to use correctly than twenty generic endpoints. Fewer ambiguous options mean fewer wrong calls and less context consumed in each session.

Who it is useful for

For Referna members who already work with an assistant such as Claude Desktop or Claude Code, the benefit is immediate: asking the agent to prepare a meeting with a contact based on what was discussed in the circle's latest sessions, without copying and pasting anything. For teams designing their own MCP servers, the announcement works as a template for communicating permissions: what is exposed, what is not and how access is revoked.

And for companies considering opening client data to external agents, the lesson is that per-user access control must live in the server itself, not depend on the good faith of the MCP client.

At ElephantPink we are seeing more and more small, well-scoped connectors like this one, and we think they are the ones that will end up building trust in MCP beyond the developer world. The missing piece is authentication: manual keys work, but an OAuth flow would be the logical next step.