flutter-mcp-toolkit-custom-tools

name	flutter-mcp-toolkit-custom-tools
description	Use this skill when the agent exposes app-specific surfaces by registering custom MCP tools and resources inside the Flutter app (mcp_toolkit dynamic registry — MCPCallEntry, bootstrapFlutter additionalEntries / addEntries). Covers tool vs resource vs evaluate-expression, Map-based handlers, schema strictness, discovery via fmt_list_client_tools_and_resources, fmt_client_tool, fmt_client_resource, and lifecycle pitfalls.
metadata	{"author":"Arenukvern"}

Custom MCP Toolkit Tools & Resources (Dynamic Registry)

Use this when bundled MCP tools (screenshot, semantic snapshot, tap, …) are not enough and you need app-specific read surfaces or actions — e.g. cart totals, feature flags, curated debug snapshots of internal state. Entries are registered in the Flutter process and exposed to the agent through the dynamic registry.

Pick the right primitive

Need	Use
One-off read of a simple value	fmt_evaluate_dart_expression (no app code change).
Stable read-only payload (diagnostics, JSON snapshot, “current route”)	MCPCallEntry.resource + fmt_client_resource. Prefer resources when the contract is “GET-like” and idempotent.
Parameterized or mutating action, or reusable named operation	MCPCallEntry.tool + fmt_client_tool.

Handler signature (tools and resources)

MCPCallHandler is FutureOr<MCPCallResult> Function(ServiceExtensionRequestMap request) where ServiceExtensionRequestMap is Map<String, String>.

• Tool arguments arrive as string values keyed by schema property names — mirror the README pattern: request['n'], request['userId'], then parse (int.tryParse, double.tryParse, jsonDecode for nested blobs if the wire format sends JSON-as-string).
• Do not use request.arguments — that is not the app-side API.

Minimal tool registration

import 'package:mcp_toolkit/mcp_toolkit.dart';

final tool = MCPCallEntry.tool(
  handler: (request) async {
    final userId = request['userId'] ?? '';
    final cart = CartRepository.instance.forUser(userId);
    return MCPCallResult(
      message: 'ok',
      parameters: {
        'total': cart.total,
        'items': cart.items.map((i) => i.toJson()).toList(),
      },
    );
  },
  definition: MCPToolDefinition(
    name: 'cart_get_snapshot',
    description: 'Return current cart total and items for a user.',
    inputSchema: {
      'type': 'object',
      'additionalProperties': false,
      'properties': {
        'userId': {'type': 'string'},
      },
      'required': ['userId'],
    },
  ),
);

await MCPToolkitBinding.instance.addEntries(entries: {tool});

Prefer MCPToolkitBinding.instance.bootstrapFlutter(additionalEntries: { ... }, runApp: ...) so tools/resources register in one place with zone/error setup — same entries shape as above.

Register after initialize() / bootstrapFlutter wiring, once at bootstrap — not inside build, not per-widget initState.

Custom resources

Resources are for read-only MCP surfaces: diagnostics, config summaries, or JSON blobs the agent polls without treating them as imperative actions.

MCPCallEntry.resource(
  definition: MCPResourceDefinition(
    name: 'app_cart_digest',
    description: 'Compact cart summary for agents (read-only).',
    mimeType: 'application/json',
  ),
  handler: (request) async => MCPCallResult(
    message: 'Cart digest',
    parameters: {
      'itemCount': CartRepository.instance.visibleCount,
      'currency': CartRepository.instance.currencyCode,
    },
  ),
),

• name must be snake_case (letters, digits, underscores). resourceUri maps it to a visual://localhost/... URI (underscore segments become path segments). Agents consume it via fmt_client_resource using that URI / listing from fmt_list_client_tools_and_resources.
• Set mimeType honestly (application/json vs text/plain) so clients know how to interpret payloads.

Schema rules (tools)

The MCP server enforces strict JSON Schema:

• Prefer additionalProperties: false unless you intentionally accept arbitrary keys. Unknown keys fail validation — good for catching agent typos.
• Mark required for anything the handler reads unconditionally.
• Prefer primitives and enum over unconstrained strings.
• parameters in MCPCallResult must be JSON-serializable; non-serializable objects degrade to toString().

Discovery from the agent side

1. fmt_list_client_tools_and_resources — enumerate app-registered tools and resources.
2. fmt_client_tool — invoke a tool by name with JSON args (CLI: flutter-mcp-toolkit exec --name fmt_client_tool --args '...' per your transport).
3. fmt_client_resource — fetch a registered resource (URI from listing / resourceUri convention).

If something should appear but does not: confirm addEntries completed (await), then hot restart — reload does not always replay discovery cleanly.

Lifecycle gotchas

• Hot reload + addEntries from widget code → duplicate registrations. Register once in main() / bootstrap, not in build.
• Hot restart clears VM state; registrations tied to bootstrapFlutter / main run again on boot — correct pattern survives restart.
• Debug mode only — release builds do not expose these VM service extensions.
• Naming: flat global namespace per app — prefix tools/resources (cart_, flags_, nav_) to avoid collisions with builtins or other domains.

When the agent authors surfaces for the user’s app

1. Ensure mcp_toolkit is in pubspec.yaml.
2. Add lib/mcp_tools/<domain>_surfaces.dart exporting registerXSurfaces() that returns Set<MCPCallEntry> or performs addEntries once.
3. Wire registerXSurfaces() from bootstrapFlutter(..., additionalEntries: ...) or call addEntries immediately after initializeFlutterToolkit inside bootstrapFlutter’s chain — never from StatefulWidget lifecycle.
4. Tight schemas (additionalProperties: false, explicit required).
5. Hot restart, then fmt_list_client_tools_and_resources before first fmt_client_tool / fmt_client_resource call.

Safety and scope

• Treat handlers as powerful debug hooks: avoid exposing secrets, full databases, or unchecked filesystem/network IO.
• Keep handlers thin: delegate to domain/services already used by the app (same DI/getters), don’t duplicate business logic in MCP-only paths unless intentional.

Common traps

• request.arguments — wrong shape; use request['key'] on Map<String, String>.
• Missing await on addEntries → race before discovery lists your surface.
• Returning Future instances inside parameters → useless serialization; await inside the handler.
• inputSchema out of sync with the handler → agents trust the schema; update both.

Related

• Driving the live app (snapshot / tap / reload): flutter-mcp-toolkit-guide → flutter-mcp-toolkit-inspect / flutter-mcp-toolkit-control.
• Repository ARCHITECTURE.md → “Dynamic Registry Architecture”.

---

Source: Arenukvern/mcp_flutter — distributed by TomeVault.