agent_kit

Drop-in Flutter widgets for the non-text parts of AI agent UIs — tool-call cards, thinking bubbles, citation chips, plan-progress timelines. Backend-agnostic. shadcn-for-agents.

The chat-UI market is crowded. The layer above chat — visualizing what the model is doing — is wide open. agent_kit fills that gap.

AgentView(stream: agent.events)

That's the entire integration. Stream AgentEvents, get a Cursor / Perplexity-style agent panel for free.

---

Install

dependencies:
  agent_kit: ^0.1.0

flutter pub get

---

30-second usage

import 'package:agent_kit/agent_kit.dart';
import 'package:flutter/material.dart';

class AnswerPanel extends StatelessWidget {
  const AnswerPanel({super.key, required this.events});

  final Stream<AgentEvent> events;

  @override
  Widget build(BuildContext context) {
    return AgentKitThemeScope(
      theme: AgentKitTheme.fromColorScheme(Theme.of(context).colorScheme),
      child: AgentView(stream: events),
    );
  }
}

Emit any of the four event kinds from your LLM client and AgentView renders the right widget for each:

controller.add(const ThinkingEvent(ThinkingStep(text: '…', isStreaming: true)));
controller.add(const ToolCallEvent(ToolCall(
  id: 'tc-1', name: 'search_web', state: ToolCallState.running,
  args: {'q': 'flutter'},
)));
controller.add(const CitationEvent([
  Citation(sourceId: 's1', title: 'flutter.dev', url: 'https://flutter.dev'),
]));
controller.add(const PlanEvent([
  PlanStep(id: 'p1', label: 'Search', state: PlanStepState.running),
  PlanStep(id: 'p2', label: 'Synthesize'),
]));

---

What you get

Widget	What it renders
ToolCallCard	Collapsible card with state icon, args (pretty-printed JSON), result preview, error + retry slot.
ToolCallFeed	Vertical feed of cards, keyed by ToolCall.id so streaming updates don't churn.
ThinkingBubble	Claude / o1-style "Thinking…" bubble with pulsing dot and auto-collapse.
ReasoningTrace	Collapsible numbered reasoning log.
CitationChip	Perplexity-style numbered inline badge.
CitationCard	Rich source card — favicon, host, title, snippet.
CitationList	Horizontal scrollable strip of citation cards.
PlanProgress	Vertical stepper with live state transitions and nested sub-steps.
AgentTimeline	Chronological mix of every event kind.
AgentView	Stream-in, widgets-out. The five-line integration.
AgentEventBuilder	Low-level builder hook for full layout control.

Every widget honors AgentKitTheme and ships with light + dark defaults derived from the ambient Material 3 ColorScheme.

---

Why agent_kit

	agent_kit	flutter_ai_toolkit	Rolling your own
Tool-call lifecycle UI	✅	❌	weeks
Thinking / reasoning UX	✅	❌	days
Citation widgets	✅	❌	days
Plan / step progress	✅	❌	days
Backend-agnostic	✅	✅	n/a
Stable widget keys (no element churn)	✅	partial	manual
Material 3 theming	✅	✅	manual
Streaming-safe state machine	✅	n/a	manual

flutter_ai_toolkit covers the chat surface (messages, typing indicators, attachments). agent_kit covers the surface above chat — agent introspection. The two compose cleanly.

---

Custom tool renderers

Every other widget is themed; for the long tail of tool calls (image generation, maps, charts) register a custom renderer once:

ToolRenderer.register('generate_image', (context, call) {
  final url = call.result as String?;
  return url == null
      ? const SizedBox.shrink()
      : Image.network(url);
});

ToolCallCard uses the registered builder when call.name matches and falls back to its default args/result preview otherwise.

---

Recipes

Use with streamdown

The two packages were designed to compose. Stream the final markdown answer with streamdown below your AgentView:

Column(
  children: [
    AgentView(stream: agent.events),
    Streamdown(stream: agent.tokens),
  ],
)

Bring your own state

Use AgentEventBuilder instead of AgentView when you need a custom layout (sidebar + main, two-column, etc.):

AgentEventBuilder(
  stream: agent.events,
  builder: (context, events) => Row(
    children: [
      Expanded(child: AnswerColumn(events: events)),
      Sidebar(events: events),
    ],
  ),
)

Wire to MCP

agent_kit's AgentEvent types map 1:1 to MCP tool-call and reasoning primitives. The companion package mcp_client (ships Sep 2026) wires this end-to-end.

---

Example

A full Cursor-style replay lives in example/ — run it on iOS, Android, web, macOS, Windows, or Linux:

cd example
flutter run

---

Quality

• 55 unit + widget tests, all green
• flutter analyze with strict casts / inference / raw-types — zero issues
• Zero runtime dependencies on AI provider SDKs
• Stable widget keys per event id — no element churn on streaming updates
• RepaintBoundary on every card

---

License

BSD-3-Clause. Copyright © 2026 Jayu Limbani.