prompt-gui-mcp

prompt-gui-mcp lets AI coding agents ask for your input through beautiful, interactive GUI forms instead of plain text prompts.

Instead of forcing an agent to guess, stall, or ask you in chat, the agent can call an MCP tool that opens a macOS desktop prompt. You answer in the app, and the result goes back to the agent as structured data.

For example, send this prompt to your agent:

Use prompt-gui-mcp to show me a form with questions that will help you recommend a travel destination.
Ask about my preferences, travel season, budget, trip style, and any other details you need.

opens a desktop form like this:

<img src="docs/assets/prompt-form-travel-example.png" alt="Travel preference prompt form" width="320">

flowchart LR
  Agent["AI coding agent"] -->|Calls prompt-form| MCP["prompt-gui-mcp"]
  MCP -->|Opens GUI form| App["Desktop app"]
  App -->|Shows prompt| User["You"]
  User -->|Submits answers| App
  App -->|Returns structured data| MCP
  MCP -->|Tool result| Agent

What It Does

• Shows flexible forms generated from an MCP tool call.
• Agents can compose forms with elements such as text, textarea, radio, select, checkbox-list, markdown, and image fields.
• Returns the user's validated answers to the calling agent.
• Keeps the prompt in a small always-on-top desktop window.
• Includes a follow-up wait tool so agents can continue waiting when a prompt takes longer than their normal tool timeout.

The current MCP tools are:

Tool	Purpose
prompt-form	Show a structured form and return submitted values plus optional feedback.
wait-for-prompt	Continue waiting for a pending prompt by UUID.

Download

Only the macOS app is packaged right now.

1. Open the GitHub Releases page.
2. Download the latest macOS .dmg or .zip asset.
3. Install and launch prompt-gui-mcp.
4. Keep the app running while your coding agent uses the MCP server.

The mac app starts the local backend sidecar automatically. By default, the MCP endpoint is:

http://127.0.0.1:43118/mcp

Set Up Your Coding Agent

Configure any MCP client that supports Streamable HTTP to connect to the local server:

{
  "mcpServers": {
    "prompt-gui-mcp": {
      "url": "http://127.0.0.1:43118/mcp"
    }
  }
}

Then restart the agent or reload its MCP servers. The desktop app must be running before the agent calls prompt-form.

Some agents use a TOML-style MCP config instead:

[mcp_servers.prompt-gui-mcp]
url = "http://127.0.0.1:43118/mcp"

After setup, ask the agent to use prompt-gui-mcp when it needs your input. For example:

Use prompt-gui-mcp to show me a form before choosing the deployment strategy. Ask for the target environment, risk tolerance, rollback preference, and approval notes.

Build And Run The Packaged App

Requirements:

• Node.js 22+
• pnpm 10+
• Rust toolchain
• Xcode Command Line Tools on macOS

Install dependencies:

pnpm install

Build the packaged desktop app:

pnpm --filter desktop tauri:build

Open the generated app bundle:

open apps/desktop/src-tauri/target/release/bundle/macos/prompt-gui-mcp.app

The .app, .dmg, and .zip outputs are written under:

apps/desktop/src-tauri/target/release/bundle/

Development Checks

pnpm --filter backend check
pnpm --filter desktop check
pnpm simulate

pnpm simulate starts the desktop app and sends a sample MCP tool call so you can test the full prompt flow.

Architecture

Coding agent -> MCP HTTP endpoint -> backend queue -> desktop app -> human
                                                 ^                    |
                                                 |--------------------|

• apps/backend is the Node.js MCP server and local HTTP API.
• apps/desktop is the Svelte frontend and Tauri v2 macOS shell.
• The backend listens on 127.0.0.1:43118 unless I_AM_MCP_SERVER_PORT is set.
• The desktop app reads backend state over local HTTP/SSE and submits completed form results back to the backend.

Repository

apps/backend   MCP server, task queue, local HTTP/SSE API
apps/desktop   Svelte UI and Tauri desktop shell
docs           Design notes, packaging notes, and README assets

Contributions are welcome. Keep changes focused, run the checks above, and open a pull request with a clear description.