Skip to main content
ODAI is built on a hub-and-spoke multi-agent architecture. A central orchestrator receives every user request and delegates work to a roster of specialized agents, each responsible for a distinct domain of capabilities. 
User Request → WebSocket → Orchestrator → Specialized Agent(s) → Response → User
Each spoke agent owns its own tools, prompts, and API credentials. The orchestrator never calls external APIs directly — it hands off to the agent best suited to fulfill the request.

​
Agent integration pattern

Every agent in ODAI follows the same structural pattern: a set of functions decorated with @function_tool and an Agent instance that exposes them. 
@function_tool(is_enabled=enable_check_function)
def agent_function(wrapper: RunContextWrapper[ChatContext], ...params) -> dict:
    """Tool description and usage instructions."""
    # API call logic
    return ToolResponse(
        response_type="agent_specific_type",
        agent_name="Service Name",
        friendly_name="Human Readable Name",
        display_response=True,
        response=processed_data
    ).to_dict()

AGENT = Agent(
    name="Agent Name",
    instructions=PROMPT_PREFIX + specific_instructions,
    handoffs=[related_agents],
    tools=[agent_function, other_tools]
)

​
@function_tool

The @function_tool decorator registers a function as a callable tool within the OpenAI Agents framework. The optional is_enabled parameter accepts a callable that receives the current RunContextWrapper and returns a boolean — this lets agents gate tools based on user permissions (e.g. whether Google OAuth or Plaid is connected).

​
ToolResponse

Every tool function returns a ToolResponse serialized to a dict. The key fields are:
FieldDescription
response_typeA string identifier for the type of data returned (e.g. "yelp_search")
agent_nameInternal agent identifier
friendly_nameHuman-readable label shown in the UI
display_responseWhen False, the tool output is not forwarded to the WebSocket client
responseThe actual payload data

​
Agent categories

ODAI includes 35+ agents organized into eight categories.

Communication & Productivity

  • Gmail Agent — send, receive, search, and reply to email
  • Google Calendar Agent — create and manage calendar events
  • Google Docs Agent — create and search documents
  • Slack Agent — team messaging integration
  • Twilio Assistant — voice and SMS capabilities

Information & Search

  • Google Search Agent — web search via SerpAPI
  • Google News Agent — top headlines and topic search
  • Fetch Website Agent — extract content from URLs
  • Google Shopping Agent — product search and price comparison

Financial Services

  • Plaid Agent — bank account balances and transactions
  • Plaid Connector Agent — guided account connection flow
  • FinnHub Agent — stock prices and financial statements
  • CoinMarketCap Agent — cryptocurrency prices and market data
  • Exchange Rate Agent — currency conversion

Travel & Transportation

  • FlightAware Agent — real-time flight status and tracking
  • Flights Agent — IATA code lookups and airport info
  • AMADEUS Agent — flight and hotel search
  • Amtrak Agent — train status and schedules
  • Caltrain Agent — regional Bay Area train information

Local Services & Entertainment

  • Yelp Agent — restaurant and business search with reviews
  • TripAdvisor Agent — travel recommendations and reviews
  • Ticketmaster Agent — event tickets and venue details
  • MovieGlu Agent — showtimes and theater search

Weather & Location

  • AccuWeather Agent — hourly and daily forecasts by coordinates
  • WeatherAPI Agent — current conditions and multi-day forecasts
  • Location Service — IP-based location detection

E-commerce & Shopping

  • Amazon Agent — product search and detail lookup
  • Google Shopping Agent — cross-retailer product comparison

Utilities & Tools

  • EasyPost Agent — package tracking across carriers
  • Open External URL Agent — open URLs in a browser window or tab
  • Google Connections Agent — Google account OAuth setup

​
Agent registration

Adding a new agent to ODAI involves four steps:
1

Implement the agent file

Create a connector file in connectors/ with @function_tool-decorated functions and an Agent instance exported as a module-level constant (e.g. MY_AGENT).
2

Import into the orchestrator

Add the import and include the agent in the handoffs list in connectors/orchestrator.py.
3

Map tool call messages

Add each tool function name to the TOOL_CALLS dict in connectors/orchestrator.py with a user-facing progress string (e.g. "Searching Yelp..."). Set the value to None for internal tools that should not surface a progress message.
4

Add integration configuration

Register the agent in integrations.yaml with an id, name, description, logo, and example prompts. This drives the frontend integrations panel and user onboarding copy.

​
Voice agent variants

Several agents expose a REALTIME_* variant alongside their standard version. These variants are registered in connectors/voice_orchestrator.py and are optimized for real-time speech: concise response formatting, no markdown, and faster execution. See Voice Interface for details.

Build docs developers (and LLMs) love

Get started for freeTalk to us