Reference: https://developers.google.com/maps/ai/agentic-ui-toolkit
Four pre-built chat UI components — not an agent framework:
| Component | What it does |
|---|---|
| Place Detail (Compact) | POI card with photo, rating, hours, address embedded in a chat bubble |
| Inline Map | Small localization map embedded mid-conversation |
| Inline Map + Route | Turn-by-turn journey preview as a self-contained chat card |
| Inline Map Detail | Rich imagery + location atmosphere card |
Plus tool-call schemas for each component so the LLM knows when to trigger them. Available on Web, Android, iOS. LLM-agnostic (Gemini / OpenAI / Anthropic compatible).
The pitch: drop these components into an AI assistant chat UI and the agent can surface rich map visuals without you building them.
Infrastructure packages:
| Package | What it covers |
|---|---|
@mapbox/agents-tools |
Tool/function calling, ToolOrchestrator, multi-agent (AgentTool), OpenTelemetry |
@mapbox/agents-conversation-manager |
Multi-turn state, ConversationManager, InMemoryStateStore |
@mapbox/agents-sqlite-state-store |
SQLite-backed persistence |
@mapbox/agents-llm-providers |
AnthropicProvider, model adapters |
@mapbox/agents-intent-matcher |
RAG-based intent → PromptTemplate routing |
@mapbox/agents-embedding-providers |
OpenAI / other embedding adapters |
@mapbox/agents-vector-db-adapters |
Qdrant adapter |
@mapbox/agents-rate-limiter |
API call rate limiting |
@mapbox/agents-ui |
Framework-agnostic chat client + React (useMapboxAgent) / Vue / Svelte adapters |
@mapbox/agents-ag-ui |
AG-UI protocol adapter (streams ToolOrchestrator as AG-UI events) |
@mapbox/agents-vercel-ai |
Vercel AI SDK adapter + Mapbox MCP bridge |
@mapbox/agents-telemetry-exporters |
Console, OTLP, Datadog exporters |
UI components (in examples):
- Atlas agent:
ChatPanel,MapPanel(GL JS, dot markers, isochrones, route display),CustomerContextCard,TracePanel - Penny concierge:
MapPanel,ItineraryPanel, full streaming chat UI; tools for search, directions, neighborhood discovery, day-plan clustering, flights (Duffel), hotels
| Google has | We have | Gap |
|---|---|---|
| Pre-built POI card (photo, rating, hours) in chat bubble | None — raw text or custom map pins | ❌ Reusable POI card component in @mapbox/agents-ui |
| Inline map widget embedded mid-conversation | MapPanel as a side panel, not in the chat stream |
❌ Inline/embedded map in chat messages |
| Route preview as a self-contained chat card | Route drawn on MapPanel |
❌ Self-contained route card component |
| Tool schema → UI component binding layer | Manual: agent emits JSON, front-end reads and updates map | ❌ Declarative schema-to-component binding |
| Android / iOS SDKs | Web only | ❌ Native mobile SDKs |
| Google Maps data grounding | Mapbox Search / Directions / Isochrone via MCP | ✅ Equivalent, different data source |
| — | Full agent infrastructure: orchestration, multi-agent, conversation state, RAG intent matching, rate limiting, telemetry, evals framework, AG-UI streaming, Vercel AI adapter | ✅ They have none of this |
Google's toolkit is purely UI widgets sitting on top of someone else's agent. We have the full backend stack — but no polished inline map UI components that render directly inside chat bubbles.
The concrete thing to build: a @mapbox/agents-ui map widget — a self-contained React component that renders a small inline map, POI card, or route preview directly inside a chat message, driven by structured tool output. That closes the only meaningful gap, and we'd have the complete picture Google is describing while they still have none of the backend.