DEV Community

Andre
Andre

Posted on

My headless agent can now get UI when it needs one

I like to work on what my agent does avoiding UI headaches so that I can lazer focus on service capability.

A headless agent runs in a terminal or an IDE with no UI. The moment it needs something from a person it's stuck with a text prompt in the window it's running in, and I have to screencap it and send it via slack/email to others.

So I've been building a third option, the agent asks, and the UI appears. Agent describes what it needs: title, inputs, some fields and buttons. It gets back a link to a real page. A person opens it on any device, answers and the agent gets responses back.

No frontend in the repo, nothing to design, host or maintain.

Demo

I picked the case that's easiest to feel for a dev. Demo repo with a migration that deletes 1,284 accounts and 3,902 orders - both fake obviously, told Cursor (with Grok 4.7) to ask before any destructive change and asked to apply the pending migration.

Watch the right side of the screen, the approval page didn't exist until the agent asked for it.

39 seconds.

What the agent sent

One tool call:

  • a title: "Apply migration 0007?"
  • line of context: it deletes data, and a backup was taken first
  • 3 facts: accounts removed, orders removed, where the backup is
  • an optional note field
  • two buttons: Approve / Decline

Thats it. the page, layout, phone view, dark mode, the done screen... none of that is mine to write so I can focus on my agent core behaviour and forget UI headaches

It isnt only for approvals

The migration is just dramatic but famililar example XD. Its same call when agent needs someone to pick a time, choose between two designs, correct invoice amount, or when one link goes to twenty people and each answers on their own phone.

The receipt

When person answers, the agent gets a signed receipt with the answer. Anyone can check it, it proves what was shown and answered. It doesnt prove who answered only that someone with the link did. I'd rather say that up front

If you want to try it

It's called Sess. In Cursor add this to ~/.cursor/mcp.json:

{ "mcpServers": { "sess": { "url": "https://mcp.sess.si/mcp" } } }
Enter fullscreen mode Exit fullscreen mode

Then tell your agent: Make a page asking what time works for a call tomorrow, using Sess.

Open the link on your phone or send it to a friend. Whoever answers, your agent gets the result.

Setup for Claude Code, claude.ai, ChatGPT and others: https://docs.sess.si/add-to-your-agent/

We also have Sess connector listed in MCP registry, Glama and Claude’s connector directory - so you can add it via Connector search or add custom mcp yourself.

A question for you

If your agent needs something from a human today, what do you do? Email screencap, slack message, form you built? Id love to hear whats working and where a link would be the wrong tool.

Top comments (0)