How it works

Your harness talks to the DreamUI MCP adapter over stdio. The adapter signs in to DreamUI cloud, runs generation, and hands back a project the user can open and keep editing.

The flow

Your harness
01
Describe
The user asks for a website or interface in Claude Code, Codex, Cursor, Muse, or any MCP harness.
DreamUI MCP
02
Generate
The harness calls uicopilot_agent_run. The adapter sends the job to DreamUI cloud and returns the result.
DreamUI
03
Refine
The user previews the result, refines it with more prompts, and keeps editing in the DreamUI workspace.

Prompt → tool call → editable website.

Architecture

There are three pieces, and only one of them is on the network:

Harness
Claude Code, Codex, Cursor, Windsurf, VS Code, Muse, or any MCP client. It decides when to call a tool.
MCP adapter
A child process the harness launches on the same machine. It speaks MCP over stdio and stores the user’s DreamUI credential locally.
DreamUI cloud
https://dreamui.com. Runs generation, stores projects and sessions, and serves the editor.

Step by step

The harness launches the adapter
When the harness starts, it spawns the adapter and calls MCP tools/list to learn the available tools and their input schemas.
The user signs in once
The agent calls uicopilot_auth. The user approves in a browser, and the adapter stores the credential on the host. See Device login.
The user describes what they want
For example: “a landing page for my pottery studio, with class booking.”
The agent calls uicopilot_agent_run
The adapter sends the prompt to DreamUI cloud and returns the generated output. With writeFiles: true it also writes files to localDir on the host.
The user keeps editing in DreamUI
The project opens in DreamUI, where the user previews, refines with more prompts, and edits directly.

Long-running work

For managed sessions, a dispatched job returns a receipt first. The agent checks progress with uicopilot_commands and reads the final state with uicopilot_session_snapshot.