lyr3

Do We Need Embeddable UIs in MCP?

•4 min read
Do We Need Embeddable UIs in MCP?

Chat boxes have become the face of AI. We type, the AI replies, and when tools are connected through MCP (Model Context Protocol), the outputs are poured back into text. Clean, minimal plumbing.

But MCP-Universe, a benchmarking project, shows just how brittle that model is. Only about one-third of complex interactions succeed when the AI has to juggle multi-step orchestration entirely through chat. That raises a question: are we overloading the AI, and forcing the chat box to do too much?

Not every task is best suited to plain text. Some are inherently visual or structured. And that’s where an embeddable UI layer could open new possibilities.

MCP Today: Text as the Only Interface

The MCP contract is simple: JSON-RPC messages define tools and resources, the AI chooses which to call, and results come back as structured data. The AI then “translates” that into chat.

  • Lookup a CRM record → AI explains the fields in text.
  • Run a query → AI describes the results row by row.
  • Generate a dashboard → AI tries to summarize trends verbally.

It works, but it’s clunky. The protocol assumes the chat box is the only place humans will consume results.

What MCP-Universe Tells Us

MCP-Universe shows how fragile this becomes as workflows grow complex. Models fail when asked to:

  • Keep track of state across multi-step queries.
  • Navigate deep hierarchies of data.
  • Translate structured results into natural language without losing context.

That’s not surprising - language is great for reasoning, but weak at precision display. A CRM dashboard, for example, is far easier to parse visually than if the AI dumps 500 words of description into a chat bubble.

Embeds as a Bridge

This is where embeddable UIs come in. Imagine a CRM MCP server responding with not just data, but a presentation hint:

{
  "url": "https://crm-mcp.com/view/opportunity/12345",
  "presentation": "embed_v1"
}

A client that supports embeds could render that as a panel inside the chat: a table of recent opportunities, or an approval form to advance a deal stage.

The AI still orchestrates - it decides when to call the CRM tool. But instead of describing rows of pipeline data, it gives the human a UI surface built for the job. This bridges the programmatic world (CRMs, structured data, forms) with the AI world (conversation, orchestration, reasoning).

Real-World Examples

  • CRM Opportunity Review: AI says, “Here are your top five deals this quarter.” Instead of a list in text, you see an embedded table with columns for account, stage, owner, and value.
  • Approval Workflow: AI: “Looks like the deal with Acme Corp is ready for approval.” Inline, an embedded card shows the deal summary with an Approve/Reject button.
  • Customer Record Update: AI: “You mentioned updating contact info. Here’s the form.” Instead of dictating JSON, you fill in a simple embedded form.

These aren’t generic iframes - they’re scoped, single-purpose UIs tied to specific MCP tools.

Benefits of an Embed Convention

  • Lighter AI load - the model doesn’t need to “speak spreadsheet” or narrate dashboards.
  • Better AIUX - humans stay in flow, working directly in context.
  • Bridge effect - connects structured programmatic systems to conversational AI in a usable way.
  • Safe scoping - embeds can be constrained to dataset views, forms, and dashboards.

Challenges & Risks

  • Security - any embed model must sandbox aggressively.
  • Spec creep - MCP must stay lean, with embeds as optional hints, not core requirements.
  • Standardization - “embed_v1” could start as convention before becoming spec.

The Way Forward

For now, presentation belongs to clients. But conventions matter. Just as OpenGraph made link previews universal, an MCP embed hint could become a quiet standard.

Not all tasks should be forced through the chat box, and not all tasks should rely on the AI alone. By signaling embeddable UIs, MCP could become the bridge between how machines store and expose data and how humans actually work with it.

The result isn’t replacing chat — it’s complementing it. Some jobs are for conversation, others for dashboards and forms. Embeds let MCP give us both.
```

← Back to Blog