Agent UI for developers

Let AI build the app. Open the code whenever you want. Every app ships with its own MCP server, and the CLI lets you pull it into your editor, work with your coding assistant and publish it back.

npm install -g @agentuiai/cli

Requires Node.js 22 or later.

$ agentui auth login
✓ Signed in as you@acme.com
$ agentui project sync <workflow-id>
✓ entities/ pages/ components/ functions/ mcp/
$ claude "add a reorder button to the stock page"
$ agentui validate --all
✓ No problems found
$ agentui project push --build
✓ Build finished · development
Works with
  • Claude Code
  • Cursor
  • Codex
  • VS Code Copilot
  • ChatGPT
  • Windsurf
Built in

Every app comes with its own MCP server.

Turn it on and Claude, ChatGPT, Cursor or any MCP client can look up and change your app's data on someone's behalf, seeing only what that person is allowed to see. No server to host, no code to write.

  • Data tools generated for every table: list, get, count, create, update and delete
  • One address per environment, off until you turn it on
  • View-only or read-and-write connections, and you can revoke any of them
  • Data Security rules checked on every call
  • Guides and custom tools that teach assistants your procedures
How to turn on MCP
  • data_orders_list
  • data_orders_get
  • data_orders_count
  • data_orders_create
  • data_orders_update
  • data_orders_delete
  • tool_summarize_overdue_invoices

Your app is just a folder

Sync an app and you get its real code: data tables as JSON, pages, components, backend functions and custom AI tools. AGENTS.md and CLAUDE.md teach your coding assistant how Agent UI apps are built.

  • Edit in any editor
  • Check it with the same validator the platform uses
  • Publish and rebuild in one command
acme-stock/
entities/ data tables as JSON
pages/
components/
functions/ → /api/<name>
mcp/ custom AI tools
Layout.jsx
AGENTS.md
.claude/CLAUDE.md

The edit loop

Change files yourself or ask your assistant to. Check, compare and publish to the environment you choose. Each environment has its own database.

  • agentui validate --all
  • agentui project diff
  • agentui project push --build
$ agentui project diff
  M pages/Stock.jsx
  A functions/reorder.js
$ agentui project push --build
✓ Uploaded 2 files
✓ Build finished · development

Connect your assistant with MCP

Agent UI's MCP server lets Claude, ChatGPT, Cursor or VS Code create apps and tables, read and update records and run your app's procedures, always with the permissions of the person who signed in. Nothing to install, no API key to copy.

Agent UI Apps

Create data-only apps, shape their tables and keep records up to date.

https://api.agentui.ai/mcp/apps

Your app

One address per environment, copied from the app's AI assistants panel. It only sees what the signed-in user may see.

How to turn it on →
$ claude mcp add --transport http agentui-apps https://api.agentui.ai/mcp/apps

Things you can ask

  • Create a data app called Field Visits with a Visits table
  • Add a status field that can only be planned, done or cancelled
  • Import these 40 rows from the spreadsheet I pasted
  • Show every open order over 500
  • Mark last month's open orders as late
  • Summarize this customer's overdue invoices

Custom MCP tools

Teach an assistant a procedure from your app, like summarizing overdue invoices or closing a ticket and notifying the owner. Ask for it in the app chat and Agent UI writes the tool as a file in your app's mcp folder.

  • Runs as the signed-in app user
  • Data Security rules still apply
  • Switch each tool on or off
  • Agent UI writes the file for you
import { defineTool } from "@/mcp";
import { Invoice } from "@/entities/Invoice";

export default defineTool({
  description: "Summarize a customer's overdue invoices",
  input: { /* JSON Schema: customerId */ },
  annotations: { readOnlyHint: true },

  async run({ customerId }) {
    const rows = await Invoice.filter({
      customer_id: customerId,
      status: "overdue",
    });
    return { count: rows.length };
  },
});

Backend functions, APIs and app signals

Every file in functions/ is reachable at /api/<name>. Connect APIs Agent UI doesn't support yet, keep keys in secrets, and send messages from your running app straight into the build chat with console.human.

  • Test functions and read their logs from the CLI
  • Secret values stay on the server
  • Separate development, staging and production environments
import { Stock } from "@/entities/Stock";

export default async function (ctx) {
  const item = await Stock.get(ctx.body.id);
  await Stock.update(item.key, { reorder: true });

  return { ok: true };
}

Install the CLI and pull your first app.

npm install -g @agentuiai/cliFree and MIT licensed. Needs Node.js 22 or later.

Browse by what you need to build

Every tool below runs on the data you already have, with role-based access, a full audit trail and a real person behind it.