Version: 0.1.33

A React-based chat interface for Pydantic AI Agents built using Agent Utilities.

Built with Vercel AI SDK and designed to work with Pydantic AI's streaming chat API.

• Streaming message responses with reasoning display
• Tool call visualization with collapsible input/output
• Interactive Elicitation Forms for structured user input
• Conversation persistence via localStorage and server-side storage
• Dynamic model and tool selection
• Dark/light theme support
• Mobile-responsive sidebar
• Scheduling jobs
• Memory
• MCP Support
• Multi-model support
• Graph Activity Visualization -- real-time specialist tracking with domain routing, parallel execution status, tool calls, and expert reasoning displayed in a collapsible timeline (GraphActivity.tsx)
• Human-in-the-loop tool approval -- security-sensitive tool calls are intercepted and require explicit user permission via an inline approval card (ApprovalCard.tsx)
• Multi-modal support -- image attachments can be sent alongside text messages for visual reasoning
• Mermaid diagram rendering via Streamdown
• Agent identity display in sidebar with workspace-aware context
• Workspace management views:
  • Files -- browse and manage workspace files
  • Skills -- view and configure universal skills
  • Scheduling -- monitor and manage cron tasks
  • Configuration -- adjust agent and workspace settings
  • Knowledge -- manage knowledge base and embeddings

• AG-UI (default): Standard Vercel AI SDK streaming via /api/chat. Supports text, reasoning, tool calls, and graph sideband events. Uses the @ai-sdk/react useChat hook for real-time streaming.
• ACP (opt-in): Advanced Agent Communication Protocol via /acp/*. Provides session management, planning modes, and approval bridges. Enabled by setting VITE_ENABLE_ACP=true. Routes through the full HSM graph pipeline via create_graph_acp_app(), ensuring ACP clients benefit from specialist routing, parallel execution, circuit breakers, and verification.

The backend (agent/agent_webui/server.py) creates a FastAPI application via create_agent_web_app() that:

1. Mounts Pydantic AI's web routes for /api/chat (model selection, tool configuration, streaming)
2. Provides enhanced workspace APIs at /api/enhanced/* (file operations, chat persistence, cron monitoring, skill management)
3. Serves the built React SPA with client-side routing support via a custom SPAStaticFiles handler
4. Integrates Logfire for real-time observability
5. Uses unified specialist discovery (discover_all_specialists()) at graph bootstrap, merging MCP agents and A2A peers into a single DiscoveredSpecialist roster before graph initialization

ACP requests route through the full HSM graph pipeline, ensuring ACP clients share the same specialist routing, parallel execution, and verification logic as AG-UI and SSE clients.

graph LR
    subgraph Sources ["Discovery Sources"]
        NA["NODE_AGENTS.md<br/>(MCP Specialists)"]
        A2A["A2A_AGENTS.md<br/>(Remote Peers)"]
    end
    subgraph Unified ["Unified Discovery (a2a.py)"]
        DAL["discover_all_specialists()"]
        DS["list[DiscoveredSpecialist]"]
    end
    subgraph Graph ["Graph Bootstrap"]
        TagPrompts["tag_prompts"]
        StepDescs["get_step_descriptions()"]
        Nodes["Specialist Nodes"]
    end
    NA --> DAL
    A2A --> DAL
    DAL -->|Deduplicated by tag| DS
    DS --> TagPrompts
    DS --> StepDescs
    DS --> Nodes

Both MCP and A2A specialists are registered through the same code path. The frontend does not need to distinguish between them -- it consumes identical sideband events (specialist_enter, tools-bound, subagent_completed) regardless of specialist source.

• Server state: React Query (@tanstack/react-query) for workspace data, conversations, and configuration
• Chat state: Vercel AI SDK useChat hook for message streaming and tool execution
• Client state: React Context and local component state
• Persistence: localStorage for conversation IDs and preferences, server-side via /api/enhanced/chats

pnpm install
pnpm run dev:server  # start the Python backend (requires agent/ setup)
pnpm run dev         # start the Vite dev server

MIT