DevUI - A Sample App for Running Agents and Workflows

A lightweight, standalone sample app interface for running entities (agents/workflows) in the Microsoft Agent Framework supporting directory-based discovery, in-memory entity registration, and sample entity gallery.

[!IMPORTANT] DevUI is a sample app to help you get started with the Agent Framework. It is not intended for production use. For production, or for features beyond what is provided in this sample app, it is recommended that you build your own custom interface and API server using the Agent Framework SDK.

Quick Start

# Install
pip install agent-framework-devui --pre

You can also launch it programmatically

from agent_framework import Agent
from agent_framework.openai import OpenAIChatClient
from agent_framework.devui import serve

def get_weather(location: str) -> str:
    """Get weather for a location."""
    return f"Weather in {location}: 72°F and sunny"

# Create your agent
agent = Agent(
    name="WeatherAgent",
    client=OpenAIChatClient(),
    tools=[get_weather]
)

# Launch debug UI - that's it!
serve(entities=[agent], auto_open=True)
# → Opens browser to http://localhost:8080

In addition, if you have agents/workflows defined in a specific directory structure (see below), you can launch DevUI from the cli to discover and run them.

# Launch web UI + API server
devui ./agents --port 8080
# → Web UI: http://localhost:8080
# → API: http://localhost:8080/v1/*

DevUI is auth-enabled by default. Localhost starts with a generated development token logged at startup; pass it as Authorization: Bearer <token> for direct API calls.

When DevUI starts with no discovered entities, it displays a sample entity gallery with curated examples from the Agent Framework repository. You can download these samples, review them, and run them locally to get started quickly.

Using MCP Tools

Important: Don't use async with context managers when creating agents with MCP tools for DevUI - connections will close before execution.

# ✅ Correct - DevUI handles cleanup automatically
mcp_tool = MCPStreamableHTTPTool(url="http://localhost:8011/mcp", client=client)
agent = Agent(tools=mcp_tool)
serve(entities=[agent])

MCP tools use lazy initialization and connect automatically on first use. DevUI attempts to clean up connections on shutdown

Resource Cleanup

Register cleanup hooks to properly close credentials and resources on shutdown:

from azure.identity.aio import DefaultAzureCredential
from agent_framework import Agent
from agent_framework.openai import OpenAIChatCompletionClient
from agent_framework_devui import register_cleanup, serve

credential = DefaultAzureCredential()
client = OpenAIChatCompletionClient()
agent = Agent(name="MyAgent", client=client)

# Register cleanup hook - credential will be closed on shutdown
register_cleanup(agent, credential.close)
serve(entities=[agent])

Works with multiple resources and file-based discovery. See tests for more examples.

Directory Structure

For your agents to be discovered by the DevUI, they must be organized in a directory structure like below. Each agent/workflow must have an __init__.py that exports the required variable (agent or workflow).

Note: .env files are optional but will be automatically loaded if present in the agent/workflow directory or parent entities directory. Use them to store API keys, configuration variables, and other environment-specific settings.

agents/
├── weather_agent/
│   ├── __init__.py      # Must export: agent = Agent(...)
│   ├── agent.py
│   └── .env             # Optional: API keys, config vars
├── my_workflow/
│   ├── __init__.py      # Must export: workflow = WorkflowBuilder(start_executor=...)...
│   ├── workflow.py
│   └── .env             # Optional: environment variables
└── .env                 # Optional: shared environment variables

Importing from External Modules

If your agents import tools or utilities from sibling directories (e.g., from tools.helpers import my_tool), you must set PYTHONPATH to include the parent directory:

# Project structure:
# backend/
# ├── agents/
# │   └── my_agent/
# │       └── agent.py    # contains: from tools.helpers import my_tool
# └── tools/
#     └── helpers.py

# Run from project root with PYTHONPATH
cd backend
PYTHONPATH=. devui ./agents --port 8080

Without PYTHONPATH, Python cannot find modules in sibling directories and DevUI will report an import error.

Viewing Telemetry (Otel Traces) in DevUI

Agent Framework emits OpenTelemetry (Otel) traces for various operations. You can view these traces in DevUI by enabling instrumentation when starting the server.

devui ./agents --instrumentation

OpenAI-Compatible API

For convenience, DevUI provides an OpenAI Responses backend API. This means you can run the backend and also use the OpenAI client sdk to connect to it. Use agent/workflow name as the entity_id in metadata, and set streaming to True as needed.

# Simple - use your entity name as the entity_id in metadata
curl -X POST http://localhost:8080/v1/responses \
  -H "Authorization: Bearer <devui-token>" \
  -H "Content-Type: application/json" \
  -d @- << 'EOF'
{
  "metadata": {"entity_id": "weather_agent"},
  "input": "Hello world"
}
EOF

Or use the OpenAI Python SDK:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8080/v1",
    api_key="<devui-token>"
)

response = client.responses.create(
    metadata={"entity_id": "weather_agent"},  # Your agent/workflow name
    input="What's the weather in Seattle?"
)

# Extract text from response
print(response.output[0].content[0].text)
# Supports streaming with stream=True

Multi-turn Conversations

Use the standard OpenAI conversation parameter for multi-turn conversations:

# Create a conversation
conversation = client.conversations.create(
    metadata={"agent_id": "weather_agent"}
)

# Use it across multiple turns
response1 = client.responses.create(
    metadata={"entity_id": "weather_agent"},
    input="What's the weather in Seattle?",
    conversation=conversation.id
)

response2 = client.responses.create(
    metadata={"entity_id": "weather_agent"},
    input="How about tomorrow?",
    conversation=conversation.id  # Continues the conversation!
)

How it works: DevUI automatically retrieves the conversation's message history from the stored thread and passes it to the agent. You don't need to manually manage message history - just provide the same conversation ID for follow-up requests.

OpenAI Proxy Mode

DevUI provides an OpenAI Proxy feature for testing OpenAI models directly through the interface without creating custom agents. Enable via Settings → OpenAI Proxy tab.

How it works: The UI sends requests to the DevUI backend (with X-Proxy-Backend: openai header), which then proxies them to OpenAI's Responses API (and Conversations API for multi-turn chats). This proxy approach keeps your OPENAI_API_KEY secure on the server—never exposed in the browser or client-side code.

Example:

curl -X POST http://localhost:8080/v1/responses \
  -H "Authorization: Bearer <devui-token>" \
  -H "X-Proxy-Backend: openai" \
  -d '{"model": "gpt-4.1-mini", "input": "Hello"}'

Note: Requires OPENAI_API_KEY environment variable configured on the backend.

CLI Options

devui [directory] [options]

Options:
  --port, -p      Port (default: 8080)
  --host          Host (default: 127.0.0.1; non-loopback hosts require auth)
  --headless      API only, no UI
  --no-open       Don't automatically open browser
  --instrumentation  Enable OpenTelemetry instrumentation
  --reload        Enable auto-reload
  --mode          developer|user (default: developer)
  --no-auth       Disable auth for loopback-only local development
  --auth-token    Custom authentication token (required for non-loopback hosts unless DEVUI_AUTH_TOKEN is set)

UI Modes

• developer (default): Full access - debug panel, entity details, hot reload, deployment
• user: Simplified UI with restricted APIs - only chat and conversation management

# Development
devui ./agents

# Local-only no-auth development
devui ./agents --no-auth

Key Endpoints

API Mapping

Given that DevUI offers an OpenAI Responses API, it internally maps messages and events from Agent Framework to OpenAI Responses API events (in _mapper.py). For transparency, this mapping is shown below:

*Uses standard OpenAI event structure but carries DevUI-specific ExecutorActionItem payload

• OpenAI = Standard OpenAI Responses API event types
• DevUI = Custom event types specific to Agent Framework (e.g., workflows, traces, function approvals)

OpenAI Responses API Compliance

DevUI follows the OpenAI Responses API specification for maximum compatibility:

OpenAI Standard Event Types Used:

• ResponseOutputItemAddedEvent - Output item notifications (function calls, images, files, data)
• ResponseOutputItemDoneEvent - Output item completion notifications
• Response.usage - Token usage (in final response, not streamed)

Custom DevUI Extensions:

• response.output_item.added with custom item types:
  • ResponseOutputImage - Agent-generated images (inline display)
  • ResponseOutputFile - Agent-generated files (inline display)
  • ResponseOutputData - Agent-generated structured data (inline display)
• response.function_approval.requested - Function approval requests (for interactive approval workflows)
• response.function_approval.responded - Function approval responses (user approval/rejection)
• response.function_result.complete - Server-side function execution results
• response.workflow_event.complete - Agent Framework workflow events
• response.trace.complete - Execution traces and internal content (DataContent, UriContent, hosted files/stores)

These custom extensions are clearly namespaced and can be safely ignored by standard OpenAI clients. Note that DevUI also uses standard OpenAI events with custom payloads (e.g., ExecutorActionItem within response.output_item.added).

Entity Management

• GET /v1/entities - List discovered agents/workflows
• GET /v1/entities/{entity_id}/info - Get detailed entity information
• POST /v1/entities/{entity_id}/reload - Hot reload entity (for development)

Execution (OpenAI Responses API)

• POST /v1/responses - Execute agent/workflow (streaming or sync)

Conversations (OpenAI Standard)

• POST /v1/conversations - Create conversation
• GET /v1/conversations/{id} - Get conversation
• POST /v1/conversations/{id} - Update conversation metadata
• DELETE /v1/conversations/{id} - Delete conversation
• GET /v1/conversations?agent_id={id} - List conversations (DevUI extension)
• POST /v1/conversations/{id}/items - Add items to conversation
• GET /v1/conversations/{id}/items - List conversation items
• GET /v1/conversations/{id}/items/{item_id} - Get conversation item

Health

• GET /health - Health check

Security

DevUI is designed as a sample application for local development and is not intended for production use. For production, or for features beyond this sample app, build a custom interface and API server using the Agent Framework SDK.

Auth is enabled by default. Unauthenticated mode is allowed only when DevUI is bound to localhost or 127.0.0.1. Network-reachable binds such as 0.0.0.0, LAN IPs, and hostnames require Bearer token authentication with an explicit token.

For shared development hosts:

# Set a token explicitly before binding beyond loopback
DEVUI_AUTH_TOKEN="<secure-dev-token>" devui ./agents --mode user --host 0.0.0.0

# Or pass the token on the command line
devui ./agents --mode user --host 0.0.0.0 --auth-token "<secure-dev-token>"

Do not use --no-auth with 0.0.0.0, LAN IPs, or hostnames. That configuration fails closed before startup.

Security features:

• User mode restricts developer-facing APIs
• Bearer token authentication is enabled by default
• Unauthenticated mode is loopback-only (localhost / 127.0.0.1)
• Non-loopback binds require DEVUI_AUTH_TOKEN or --auth-token
• Only loads entities from local directories or in-memory registration
• No remote code execution capabilities
• Binds to localhost (127.0.0.1) by default

Best practices:

• Do not use DevUI as a production deployment surface
• Use --mode user plus DEVUI_AUTH_TOKEN or --auth-token for shared development hosts
• Review all agent/workflow code before running
• Only load entities from trusted sources
• Use .env files for sensitive credentials (never commit them)

Implementation

• Discovery: agent_framework_devui/_discovery.py
• Execution: agent_framework_devui/_executor.py
• Message Mapping: agent_framework_devui/_mapper.py
• Conversations: agent_framework_devui/_conversations.py
• API Server: agent_framework_devui/_server.py
• CLI: agent_framework_devui/_cli.py

Examples

See working implementations in python/samples/02-agents/devui/

License

MIT