Flutter MCP

Flutter MCP

Golden path for agents driving a live Flutter app via the flutter-mcp-toolkit MCP server entry. Older configs may use the legacy flutter-inspector mcpServers registry id for the same binary (that id is not the Claude subagent flutter-mcp-toolkit-runtime).

When to use

• User references a running Flutter app (debug mode) and wants to inspect, screenshot, interact with, or hot-reload it.
• User pastes a VM service URI (ws://127.0.0.1:8181/.../ws) or mentions port 8181.
• You need runtime proof (before/after screenshots) that a code edit took effect.

Preflight (always first)

Before any VM-dependent call:

1. Call doctor (or run flutter-mcp-toolkit doctor --json) — parses env, ports, app reachability.
2. Confirm required toolkit extensions exist on the target:
   • ext.mcp.toolkit.app_errors
   • ext.mcp.toolkit.view_details
   • ext.mcp.toolkit.view_screenshots
   • ext.mcp.toolkit.inspect_widget_at_point

If missing, stop and report the instrumentation gap — do not guess:

• Add mcp_toolkit to pubspec.yaml.
• Ensure MCPToolkitBinding.instance..initialize()..initializeFlutterToolkit(); runs before runApp.
• Hot restart (not reload) — reload is often insufficient for extension registration.

Tool naming (v3.0.0+)

All MCP tools surface under the fmt_ capability prefix (fmt_tap_widget, fmt_hot_reload_and_capture, etc.). The prefix is mandatory in tools/call. Skill-local references below use the bare name for readability — when invoking, prepend fmt_. Dynamic-registry host tools (fmt_list_client_tools_and_resources, fmt_client_tool, fmt_client_resource) use the same prefix for a single consistent surface.

Interaction loop (Playwright-style)

1. fmt_semantic_snapshot → returns s_0..s_N refs + snapshot_id.
2. fmt_tap_widget / fmt_enter_text / fmt_scroll / fmt_swipe / fmt_long_press / fmt_drag — act on a ref. Always pass snapshotId — you get a structured stale_snapshot error if the tree moved, instead of a silent wrong tap.
3. fmt_reveal_search — for off-screen semantic targets, bounded snapshot/search/scroll that returns a fresh ref + snapshotId.
4. fmt_evaluate_dart_expression — read state directly (e.g. AgentState.instance.counter).
5. fmt_hot_reload_and_capture — after a code edit, returns reload status + screenshot + fresh snapshot + errors in one response. Prefer this over manual reload + separate capture.

Error envelope contract

Errors are {code, message, details, descriptor, recovery}. Parse error.descriptor (not the top-level object) for the machine-readable shape. Strict schemas default to additionalProperties: false — unknown params reject.

Common codes and recovery:

• connection_selection_required — retry with arguments.connection.targetId or exact arguments.connection.uri from app.debugPort.wsUri.
• target_not_found — refresh targets, then prefer exact arguments.connection.uri.
• stale_snapshot — call fmt_semantic_snapshot again, then retry the action with the new snapshotId input.
• tool_not_found — confirm the prefixed name (fmt_<tool>); v3.0.0 dropped legacy unprefixed names.
• Empty screenshot output — verify the server was not started with --no-images.
• Missing view resource/tool — verify the server was not started with --no-resources.

Visual QA

• Before/after screenshots are the proof artifact for any UI claim. Capture before with fmt_capture_ui_snapshot (or one of the visual://localhost/... resources), edit, fmt_hot_reload_and_capture, compare.
• For each reported visual issue, attach coordinate + fmt_inspect_widget_at_point output.
• Map defects to source via fmt_get_app_errors top stack frame (file, line, column) when available.
• Do not use fmt_debug_dump_* unless explicitly requested (server must be started with --dumps; high token cost).

Permissions (macOS)

Screen Recording permission belongs to the process running flutter-mcp-toolkit (or the MCP server host). If visual capture is denied:

flutter-mcp-toolkit permissions status
flutter-mcp-toolkit permissions request
flutter-mcp-toolkit permissions open-settings

Non-modifiable apps

If the target app cannot be instrumented (third-party binary, restricted env), report flutter-mcp as unavailable for that app. Do not claim screenshot/layout/error inspection success.

Related

• For the dynamic-tools side (registering custom MCP tools from inside the Flutter app), see the flutter-mcp-toolkit-custom-tools skill.
• For routing across setup / inspect / control / debug skills, see flutter-mcp-toolkit-guide.