flutter_cockpit

The first-class Flutter development adapter for Cockpit.

flutter_cockpit is Cockpit's first-class in-app adapter for AI-driven Flutter development, inspection, and control. It is independent from the black-box path used for installed production applications.

It provides:

runtime bootstrap through FlutterCockpit.runApp or FlutterCockpitApp
command execution for taps, text input, gestures, waits, assertions, screenshots, and snapshots
remote session serving over HTTP
structured Widget, Element, RenderObject, semantics, route, focus, log, runtime error, network, and rebuild state
snapshot, artifact, recording, and bundle models
target, plane, surface, and fallback-aware runtime models for AI-first summaries

Install

Requires Flutter 3.32.0 or newer.

# cockpit/pubspec.yaml
dev_dependencies:
  flutter_cockpit: any

Keep the runtime development-only. Put every flutter_cockpit import and all integration code under cockpit/; production lib/ code and production entrypoints remain unchanged.

Darwin integration supports both CocoaPods and Swift Package Manager. The package includes an iOS and macOS .podspec as well as Package.swift manifests backed by the same native sources and privacy manifests. Flutter uses the integration selected by the host project, so CocoaPods projects do not need to migrate to SwiftPM.

The runtime package declares native plugin entries for Android, iOS, macOS, Linux, Windows, and web. That lets app-window screenshots and recording fallbacks register consistently whenever the cockpit entrypoint is compiled. Keep the integration isolated by importing it only from cockpit/, never from production lib/ code. Flutter-view screenshots, Element-based inspection and control, network signals, runtime diagnostics, and remote sessions work in-app without requiring application-authored Semantics. System dialogs, notifications, host screenshots, and host recordings should still be driven by cockpit system actions so capability discovery and platform fallbacks remain truthful.

Install For AI Agents

Ask the current AI host to install the CLI, complete Skill, native adapter, and MCP surface:

First fetch and read the complete Cockpit installation guide with `curl -fsSL https://raw.githubusercontent.com/cockpit-dev/cockpit/main/skills/cockpit/INSTALL.md`, then install or update the CLI, complete cockpit Skill, native adapter, and cockpit_mcp for the current AI host exactly as that guide directs.

The guide covers Codex, Claude Code, Cursor, Gemini CLI, Kiro, OpenCode, Pi, Oh My Pi, Cline, GitHub Copilot, Windsurf, Roo Code, and portable fallback installation.

Recommended Integration

Create a non-published Flutter package under cockpit/. It depends locally on the real application and keeps flutter_cockpit in the shell's dev_dependencies; neither dependency enters the production package graph. The globally installed cockpit CLI is not an application dependency. Keep the normal production entrypoint and production lib/ untouched. Do not add flutter_cockpit imports to production lib/ code.

# cockpit/pubspec.yaml
name: your_app_cockpit
publish_to: none

environment:
  sdk: '>=3.8.0 <4.0.0'
  flutter: '>=3.32.0'

dependencies:
  flutter:
    sdk: flutter
  your_app:
    path: ..

dev_dependencies:
  flutter_cockpit: any

Replace your_app with the actual application package name and run flutter pub get inside cockpit/.

If the application uses a Pub workspace, add cockpit/ to the root workspace list, add resolution: workspace to the shell manifest, and use a compatible application version constraint instead of path: ... Run flutter pub get from the workspace root. This keeps the shell locally resolved without adding Cockpit to the production package dependencies.

import 'package:flutter/material.dart';
import 'package:flutter_cockpit/flutter_cockpit_flutter.dart';

import 'package:your_app/app_shell.dart';

Future<void> main() async {
  runApp(buildCockpitDevelopmentApp());
}

Widget buildCockpitDevelopmentApp() {
  return FlutterCockpitApp(
    config: FlutterCockpitConfig.production(
      remoteSession: CockpitRemoteSessionConfiguration.resolveFromEnvironment(
        fallback: const CockpitRemoteSessionConfiguration(
          enabled: true,
          host: '127.0.0.1',
          port: 47331,
        ),
      ),
    ),
    child: MaterialApp(
      navigatorObservers: <NavigatorObserver>[
        FlutterCockpit.navigatorObserver,
      ],
      home: const AppShell(),
    ),
  );
}

Replace package:your_app/app_shell.dart with the import that already exposes your app root widget or bootstrap. Cockpit's target launch operation injects the FLUTTER_COCKPIT_REMOTE_* dart-defines, so resolveFromEnvironment(...) enables the remote surface without taking over the production bootstrap. Only wire FlutterCockpit.navigatorObserver from the standalone shell entrypoint. FlutterCockpitApp automatically discovers the public RouteInformationProvider used by Flutter Router, RouterConfig, go_router, and other Router-based libraries, so an app-owned router normally needs no additional route bridge.

For nested navigators, create one observer per navigator so route state can return to the parent stack after a nested pop:

Navigator(
  observers: <NavigatorObserver>[
    FlutterCockpit.createNavigatorObserver(),
  ],
  onGenerateRoute: buildRoute,
)

The same factory works with router libraries that expose navigator observers, including root and shell navigators. For dynamically created routers that cannot be discovered from the mounted tree, bind their public provider from cockpit/ with FlutterCockpit.bindRouteInformationProvider(...). Use FlutterCockpit.setCurrentRouteName(...) only when a router exposes neither a provider nor observers; flutter_cockpit does not depend on any third-party router package.

Run it with:

cd cockpit
flutter run --target main.dart

What The Runtime Exposes

low-intrusion root bootstrap
command routing and execution
UI snapshots plus minimal, standard, and full mounted Element trees
accessibility, network, runtime, and rebuild signals
screenshot and recording requests
remote session status and command endpoints

HTTP diagnostics redact credential values by default while retaining useful structure such as authorization schemes, cookie names, query keys, and JSON field names. A development-only entrypoint can explicitly use CockpitHttpNetworkObserverConfiguration(redact: false) when raw bounded payloads are required; never enable raw capture in a production entrypoint or an evidence-producing CI run.

FlutterCockpitRoot treats Flutter hot reload as a runtime-diagnostic generation boundary. Errors and unconsumed recorded steps from the previous generation are cleared during reassembly; errors raised by the reloaded application are captured normally.

Host-side orchestration, MCP, workspace tooling, and delivery validation live in cockpit. The runtime bundle models now preserve targetKind, primaryExecutionPlane, planesUsed, surfaceKindsUsed, fallbackCount, plus per-step and per-observation plane metadata so host-side tooling can explain when Flutter control stayed on-plan versus when it had to degrade to another surface. On web, the runtime supports the Flutter Element and Flutter-view control path directly, while the method channels are registered as explicit unavailable stubs so capability checks stay truthful instead of failing through missing-plugin noise. On mobile and desktop, native method-channel recording and capture register through the package plugin entries and are used as app-window evidence fallbacks; prefer system or host evidence through cockpit when the goal is to prove system dialogs, notifications, host windows, or cross-app behavior.

Package page: pub.dev/packages/flutter_cockpit

flutter_cockpit

Install

Install For AI Agents

Recommended Integration

What The Runtime Exposes

Libraries

flutter_cockpit package