MCP UI Playground

A runnable lab for comparing plain MCP results, FastMCP Prefab UI, server-interactive MCP Apps, and a raw HTML MCP Apps bridge inside ChatGPT and other compatible hosts.

Run locally

uv sync --dev
uv run mcpui-playground

The Streamable HTTP endpoint is http://127.0.0.1:8775/mcp.

Inspect it with:

npx @modelcontextprotocol/inspector@latest

Or open FastMCP's local Apps preview:

uv run fastmcp dev apps src/mcpui_playground/server.py:mcp

Related MCP server: MCP Apps POC

Experiments

Preview

The local FastMCP host also exercises server callbacks and the raw MCP Apps bridge. Screenshots and Codex CLI probe outputs are indexed in artifacts/README.md.

Connect from ChatGPT web

This checkout is deployed on the machine's existing Tailscale Funnel at:

https://riema.taile42c02.ts.net/mcpui/mcp

In ChatGPT, enable Settings -> Security and login -> Developer mode, open the Plugins directory, select the plus button, and add the URL above. After tool scanning succeeds, start a new chat, enable the development plugin, and try:

Open the MCP UI Lab metrics dashboard and then open the experiment notebook.

The endpoint is intentionally unauthenticated and serves only synthetic, process-local demo data. It shares the standard HTTPS port through a /mcpui Funnel mount; the existing Jacobian /mcp route remains separate. See research notes for protocol details, constraints, and primary sources.

Pinned upstream source references are cataloged in references/README.md. Materialize them locally with ./scripts/sync-references.sh.

Use from Codex CLI

Codex CLI can call the same MCP server, but its terminal UI does not embed the MCP App iframe. Add the local service and use the terminal-native fallback. The editable scenario also supports a conversational update-and-redraw loop:

codex mcp add mcpui-playground --url http://127.0.0.1:8775/mcp
codex "Call show_terminal_dashboard from mcpui-playground and display it verbatim."
codex "Set ui_opens to 2000 with update_scenario_metric, then call show_scenario_terminal."

The Prefab tool is still useful to Codex as structured data: Codex receives and can reason over its component tree, while a compatible web host performs the interactive rendering. See docs/codex-cli-findings.md for the recorded probe.

Verify

uv run pytest
uv run ruff check .