turbo_bridge_mcp

MCP (Model Context Protocol) server that gives LLMs direct access to running Flutter apps. Connect Claude, Cursor, VS Code Copilot, or any MCP-compatible host to see, understand, and interact with Flutter UIs.

What It Does

Global install for MCP hosts

dart pub global activate turbo_bridge_mcp


Or install it per project:

```bash
dart pub add --dev turbo_bridge_mcp
An LLM connected via this server can:
- Take screenshots and see the app's current state
- Inspect the widget tree to understand UI structure
- Tap buttons and interact with the app
- Find specific widgets by text, key, or type
- Query app metadata (screen size, platform, theme)

## Setup

### Prerequisites

1. A Flutter app with `turbo_bridge` running (see [turbo_bridge](../turbo_bridge/))
2. An MCP-compatible LLM host
3. Dart SDK (comes with Flutter)

### Quick Install

```bash
# Clone the repo (one-time setup)
git clone https://github.com/mark-nicepants/flutter_turbo_bridge.git
cd flutter_turbo_bridge && dart pub get --directory=packages/turbo_bridge_mcp

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "turbo_bridge": {
      "command": "turbo_bridge_mcp"
    }
  }
}

Note: If Claude Desktop does not inherit your Dart pub cache bin path, replace turbo_bridge_mcp with the absolute executable path reported by dart pub global activate turbo_bridge_mcp.

VS Code (GitHub Copilot)

Add to .vscode/mcp.json in your project (committable to version control):

{
  "servers": {
    "turbo_bridge": {
      "command": "dart",
      "args": ["run", "turbo_bridge_mcp"]
    }
  }
}

Tip: Prefer dart run turbo_bridge_mcp for project-local installs and turbo_bridge_mcp for globally activated installs.

VS Code From Another App Workspace

If you are working in an app repo outside this monorepo and want to point at a local checkout of flutter_turbo_driver, define the MCP server with a relative or absolute path to the CLI entrypoint.

Example from an external Flutter app workspace:

{
  "servers": {
    "turbo_bridge": {
      "command": "turbo_bridge_mcp",
      "args": ["--bridge-port", "8888"]
    }
  }
}

This setup is useful when the app under test and flutter_turbo_driver live in sibling folders on the same machine and you want the MCP server definition checked into the app workspace.

Cursor

Add to .cursor/mcp.json in your project (committable):

{
  "mcpServers": {
    "turbo_bridge": {
      "command": "dart",
      "args": ["run", "turbo_bridge_mcp"]
    }
  }
}

Android Devices / Emulators

The bridge server runs inside the app on the device, so localhost:8888 on the host can't reach it directly. The MCP server automatically detects this and sets up ADB port forwarding — no extra configuration needed.

How it works:

On startup, the MCP server tries to reach the bridge at localhost:8888
If unreachable, it checks for a connected Android device via adb devices
If found, it runs adb forward tcp:8888 tcp:8888 automatically
Once the bridge is reachable, it reads /info; if DevTools is enabled, it also forwards the DevTools UI port (for example adb forward tcp:8889 tcp:8889)
On exit, it cleans up every port it forwarded

No changes to your MCP config are needed — the same setup works for both desktop and Android.

CLI Options

dart run bin/turbo_bridge_mcp.dart [options]

-h, --bridge-host    Turbo Bridge host (default: localhost)
-p, --bridge-port    Turbo Bridge port (default: 8888)
    --vm-uri         Dart VM Service URI (optional, for expression evaluation)
    --help           Show usage

MCP Tools

`screenshot`

Captures the app's current screen as a PNG image.

Parameters:

pixelRatio (number, optional) — Device pixel ratio, default 1.0
delayMs (number, optional) — Wait before capture, default 75ms

Returns: PNG image plus JSON metadata. The JSON metadata includes _meta.startedAtUtc, _meta.completedAtUtc, captureTimeMs, and roundTripMs.

`widget_tree`

Returns the full widget tree as structured JSON.

Parameters:

depth (number, optional) — Max tree depth, default 10
x (number, optional) — Focus X coordinate in logical pixels
y (number, optional) — Focus Y coordinate in logical pixels
ancestorLevels (number, optional) — Ancestors to keep above the focused hit node, default 2

Returns: Indented JSON tree with widget types, keys, text, and bounds. Responses include _meta.startedAtUtc, _meta.completedAtUtc, captureTimeMs, and roundTripMs. When x and y are provided together, the tool asks the bridge for a smaller local subtree around that coordinate.

`tap`

Injects a tap at exact screen coordinates.

Parameters:

x (number, required) — X coordinate
y (number, required) — Y coordinate

Returns: Success status plus _meta.startedAtUtc, _meta.completedAtUtc, executionTimeMs, and roundTripMs

`app_info`

Returns app metadata: screen size, pixel ratio, platform, dark mode, bridge version, and MCP compatibility hints.

Returns: App metadata plus _meta.startedAtUtc, _meta.completedAtUtc, mcpServerVersion, mcpVersionStatus, and an updateHint when the bridge is newer than the MCP server

`find_widget`

Searches the widget tree by text, key, or type. Returns matching widgets with their center coordinates for tapping.

Parameters:

text (string, optional) — Find by text content
key (string, optional) — Find by widget key
type (string, optional) — Find by widget type name

Returns: Up to 10 matching widgets with bounds and center coordinates, plus _meta.startedAtUtc, _meta.completedAtUtc, searchTimeMs, and roundTripMs

`flutter_swipe`

Performs a swipe gesture between two points.

Parameters:

startX (number, required) — Start X coordinate
startY (number, required) — Start Y coordinate
endX (number, required) — End X coordinate
endY (number, required) — End Y coordinate
steps (number, optional) — Number of move events, default 10

Returns: Success status plus _meta.startedAtUtc, _meta.completedAtUtc, executionTimeMs, and roundTripMs

`flutter_scroll`

Scrolls at a specific position.

Parameters:

x (number, required) — Scroll position X
y (number, required) — Scroll position Y
dy (number, optional) — Vertical scroll delta (positive = content up, negative = content down)
dx (number, optional) — Horizontal scroll delta

Returns: Success status plus _meta.startedAtUtc, _meta.completedAtUtc, executionTimeMs, and roundTripMs

`flutter_enter_text`

Enters text into the currently focused text field.

Parameters:

text (string, required) — Text to enter
replace (boolean, optional) — Replace existing text, default false

Returns: Success status plus _meta.startedAtUtc, _meta.completedAtUtc, executionTimeMs, and roundTripMs

`flutter_recent_logs`

Returns recent app-emitted log lines that the app pushed into TurboBridge.instance.logs. Use this to inspect what the running app has been doing without hooking into the device's system logs.

Parameters:

limit (integer, optional) — Max entries to return (default 50)
level (string, optional) — Minimum severity: trace, debug, info, warn, error

Returns: { entries: [{ id, timestamp, level, message, category?, data?, error?, stackTrace? }], count }

`flutter_recent_network`

Returns recent network calls that the app pushed into TurboBridge.instance.network from its HTTP client interceptors.

Parameters:

limit (integer, optional) — Max calls to return (default 50)

Returns: { entries: [{ id, timestamp, method, url, status, durationMs, error?, responseBodySize? }], count }

Timing Model

Every MCP tool and resource response now includes a _meta object with UTC start and completion stamps. Use it like this:

action start: _meta.startedAtUtc
action end: _meta.completedAtUtc
action wall-clock duration: difference between _meta.completedAtUtc and _meta.startedAtUtc
operation-specific timing: keep captureTimeMs, searchTimeMs, executionTimeMs, and roundTripMs as separate fields for transport and bridge analysis

For a full scenario or benchmark run, compute total wall-clock duration from the first tool call's startedAtUtc to the last tool call's completedAtUtc.

MCP Resources

`flutter://app/info`

Live app metadata (screen size, platform, theme).

`flutter://app/tree`

Current widget tree snapshot.

MCP Prompts

`flutter_inspect`

Generates a structured inspection workflow prompt. Useful for guiding an LLM through systematic app exploration.

Arguments:

focus (string, optional) — Specific area or feature to focus inspection on

Example Conversation

Once connected, you can ask the LLM things like:

"Take a screenshot of the app and tell me what you see"

"Find the login button and tap it"

"Scroll down in the list to find item 50"

"Swipe left to dismiss the notification"

"Enter 'hello@example.com' in the email field"

"Inspect the widget tree and describe the navigation structure"

"What's the current screen size and platform?"

The LLM will use the MCP tools autonomously to answer.

Architecture

LLM Host ──stdio──► turbo_bridge_mcp ──HTTP──► Flutter App
                          │                        │
                    mcp_dart server          turbo_bridge
                          │                   (in-app)
                    turbo_bridge_client
                     (Dart HTTP client)

The MCP server communicates with LLM hosts via stdio (stdin/stdout JSON-RPC) and with the Flutter app via the Turbo Bridge HTTP API. It's a thin adapter layer — all the heavy lifting happens in the Flutter app process.

turbo_bridge_mcp

What It Does

Global install for MCP hosts

Claude Desktop

VS Code (GitHub Copilot)

VS Code From Another App Workspace

Cursor

Android Devices / Emulators

CLI Options

MCP Tools

`screenshot`

`widget_tree`

`tap`

`app_info`

`find_widget`

`flutter_swipe`

`flutter_scroll`

`flutter_enter_text`

`flutter_recent_logs`

`flutter_recent_network`

Timing Model

MCP Resources

`flutter://app/info`

`flutter://app/tree`

MCP Prompts

`flutter_inspect`

Example Conversation

Architecture

Libraries

turbo_bridge_mcp package