Skip to main content

Exa

The Exa agent connector is a Python package that equips AI agents to interact with Exa through strongly typed, well-documented tools. It's ready to use directly in your Python app, in an agent framework, or exposed through an MCP.

Exa is an AI-powered search engine that finds the exact content you're looking for on the web using embeddings-based search. This connector provides access to Exa's search, contents retrieval, and find-similar endpoints. All endpoints use POST requests with JSON bodies. Requires an Exa API key from dashboard.exa.ai.

Example prompts​

The Exa connector is optimized to handle prompts like these.

  • Search for latest news on Airbyte
  • Find web pages similar to https://airbyte.com
  • Get the full text content of https://airbyte.com
  • Search for AI research papers published this year
  • Find company pages related to data integration startups

Entities and actions​

This connector supports the following entities and actions. For more details, see this connector's full reference documentation.

EntityActions
Search ResultsList
ContentsList
Similar ResultsList

Exa API docs​

See the official Exa API reference.

Interfaces​

Use the Exa connector through the Airbyte Agent CLI, the Python SDK, or the API.

CLI​

Install the CLI:

curl -fsSL https://airbyte.ai/install.sh | bash

Authenticate with Airbyte:

airbyte-agent login

Create the connector. The CLI opens the hosted setup flow:

airbyte-agent connectors create --json '{
  "workspace": "<your_workspace_name>",
  "name": "exa"
}'

Describe the connector to see its supported entities and actions:

airbyte-agent connectors describe --json '{
  "workspace": "<your_workspace_name>",
  "name": "exa"
}'

Execute an action:

airbyte-agent connectors execute --json '{
  "workspace": "<your_workspace_name>",
  "name": "exa",
  "entity": "search_results",
  "action": "list"
}'

Python SDK​

Installation​

uv pip install airbyte-agent-sdk

Usage​

Connectors can run in hosted or open source mode.

Hosted​

In hosted mode, API credentials are stored securely in Airbyte Agents. You provide your Airbyte credentials instead. If your Airbyte client can access multiple organizations, also set organization_id.

This example assumes you've already authenticated your connector with Airbyte. See Authentication to learn more about authenticating. If you need a step-by-step guide, see the hosted execution tutorial.

The connect() factory returns a fully typed ExaConnector and reads AIRBYTE_CLIENT_ID / AIRBYTE_CLIENT_SECRET from the environment:

The recommended pattern is build_connector_tools, which gives the agent three tools bound to this connector: inspect_connector, read_skill_docs, and execute. The agent can inspect the connector, read only the skill-doc section it needs, and then execute:

inspect_connector() -> read_skill_docs() -> read_skill_docs(section="...") -> execute(entity, action, params)

Pass section IDs verbatim as the outline lists them, prefix included (actions.<entity>.<action>, not <entity>.<action>); anything else returns an error the agent has to recover from.

The builder names its tools inspect_connector, read_skill_docs, and execute, so the tool sets for more than one connector collide when registered on the same agent. Renaming the callables at registration avoids the collision, but the generated execute guidance still names inspect_connector and read_skill_docs, pointing the model at the wrong tools. Use the agent_tool pattern below instead: it weaves your own names into that guidance.

Pydantic AI
from airbyte_agent_sdk import build_connector_tools
from pydantic_ai import Agent
from airbyte_agent_sdk import connect
from airbyte_agent_sdk.connectors.exa import ExaConnector

connector = connect("exa", workspace_name="<your_workspace_name>")

tools = build_connector_tools(connector, framework="pydantic_ai")
agent = Agent("openai:gpt-4o", tools=tools.as_list())
Custom tool bodies​

When you need custom tool bodies — or a framework without native support — use ExaConnector.agent_tool. Register execute, inspect, and docs together so the agent can fetch connector guidance progressively. Pass the framework explicitly when it has a supported failure strategy:

Pydantic AI
from pydantic_ai import Agent
from airbyte_agent_sdk import connect
from airbyte_agent_sdk.connectors.exa import ExaConnector

connector = connect("exa", workspace_name="<your_workspace_name>")

agent = Agent("openai:gpt-4o")

@agent.tool_plain
@ExaConnector.agent_tool(
    framework="pydantic_ai",
    inspect_tool="exa_inspect",
    docs_tool="exa_read_docs",
)
async def exa_execute(entity: str, action: str, params: dict | None = None):
    return await connector.execute(entity, action, params or {})

@agent.tool_plain
@ExaConnector.agent_tool(framework="pydantic_ai")
async def exa_inspect():
    return await connector.inspect_connector()

@agent.tool_plain
@ExaConnector.agent_tool(framework="pydantic_ai")
async def exa_read_docs(section: str | None = None):
    return await connector.read_skill_docs(section)

Use the same three-function pattern with framework="langchain", "openai_agents", or "mcp" and that framework's registration decorator. Each value translates connector failures into the framework's own signal:

framework=Tool failures surface as
"pydantic_ai"pydantic_ai.ModelRetry
"langchain"langchain_core.tools.ToolException (set handle_tool_error=True to feed it back to the model)
"openai_agents"the failure message returned to the model as the tool result
"mcp"fastmcp.exceptions.ToolError
"none" (default)airbyte_agent_sdk.AirbyteToolError

On a framework the SDK does not support natively — or in a raw LLM dispatch loop — omit framework= and handle AirbyteToolError yourself:

No framework
from airbyte_agent_sdk import AirbyteToolError
from airbyte_agent_sdk import connect
from airbyte_agent_sdk.connectors.exa import ExaConnector

connector = connect("exa", workspace_name="<your_workspace_name>")

@ExaConnector.agent_tool(
    inspect_tool="exa_inspect",
    docs_tool="exa_read_docs",
)
async def exa_execute(entity: str, action: str, params: dict | None = None):
    return await connector.execute(entity, action, params or {})

@ExaConnector.agent_tool()
async def exa_inspect():
    return await connector.inspect_connector()

@ExaConnector.agent_tool()
async def exa_read_docs(section: str | None = None):
    return await connector.read_skill_docs(section)

# Advertise all three to the model, using each function's docstring as its description.
handlers = {
    fn.__name__: fn
    for fn in (exa_inspect, exa_read_docs, exa_execute)
}

# `tool_name` and `tool_args` come from the model's tool call in your dispatch loop.
try:
    tool_result = await handlers[tool_name](**tool_args)
except AirbyteToolError as err:
    tool_result = str(err)  # hand the message back to the model as an errored tool result

Each function's docstring carries the guidance the model needs, so pass it through as the tool description wherever you register it.

Legacy alternatives​

These examples are kept for existing integrations. The deprecated ExaConnector.tool_utils pattern loads the connector's full generated catalog into one broad execute tool description instead of letting the agent read skill docs on demand. For new code, use build_connector_tools or ExaConnector.agent_tool above.

Pydantic AI
from pydantic_ai import Agent
from airbyte_agent_sdk import connect
from airbyte_agent_sdk.connectors.exa import ExaConnector

connector = connect("exa", workspace_name="<your_workspace_name>")

agent = Agent("openai:gpt-4o")

@agent.tool_plain
@ExaConnector.tool_utils
async def exa_execute(entity: str, action: str, params: dict | None = None):
    return await connector.execute(entity, action, params or {})

Or pass credentials explicitly (equivalent, useful when you're not loading them from the environment):

Pydantic AI
from airbyte_agent_sdk import build_connector_tools
from pydantic_ai import Agent
from airbyte_agent_sdk.connectors.exa import ExaConnector
from airbyte_agent_sdk.types import AirbyteAuthConfig

connector = ExaConnector(
    auth_config=AirbyteAuthConfig(
        workspace_name="<your_workspace_name>",
        organization_id="<your_organization_id>",  # Optional for multi-org clients
        airbyte_client_id="<your-client-id>",
        airbyte_client_secret="<your-client-secret>"
    )
)

tools = build_connector_tools(connector, framework="pydantic_ai")
agent = Agent("openai:gpt-4o", tools=tools.as_list())
Open source​

In open source mode, you provide API credentials directly to the connector.

The recommended pattern is build_connector_tools, which gives the agent three tools bound to this connector: inspect_connector, read_skill_docs, and execute. The agent can inspect the connector, read only the skill-doc section it needs, and then execute:

inspect_connector() -> read_skill_docs() -> read_skill_docs(section="...") -> execute(entity, action, params)

Pass section IDs verbatim as the outline lists them, prefix included (actions.<entity>.<action>, not <entity>.<action>); anything else returns an error the agent has to recover from.

The builder names its tools inspect_connector, read_skill_docs, and execute, so the tool sets for more than one connector collide when registered on the same agent. Renaming the callables at registration avoids the collision, but the generated execute guidance still names inspect_connector and read_skill_docs, pointing the model at the wrong tools. Use the agent_tool pattern below instead: it weaves your own names into that guidance.

Pydantic AI
from airbyte_agent_sdk import build_connector_tools
from pydantic_ai import Agent
from airbyte_agent_sdk.connectors.exa import ExaConnector
from airbyte_agent_sdk.connectors.exa.models import ExaAuthConfig

connector = ExaConnector(
    auth_config=ExaAuthConfig(
        api_key="<Your Exa API key from dashboard.exa.ai/api-keys>"
    )
)

tools = build_connector_tools(connector, framework="pydantic_ai")
agent = Agent("openai:gpt-4o", tools=tools.as_list())
Custom tool bodies​

When you need custom tool bodies — or a framework without native support — use ExaConnector.agent_tool. Register execute, inspect, and docs together so the agent can fetch connector guidance progressively. Pass the framework explicitly when it has a supported failure strategy:

Pydantic AI
from pydantic_ai import Agent
from airbyte_agent_sdk.connectors.exa import ExaConnector
from airbyte_agent_sdk.connectors.exa.models import ExaAuthConfig

connector = ExaConnector(
    auth_config=ExaAuthConfig(
        api_key="<Your Exa API key from dashboard.exa.ai/api-keys>"
    )
)

agent = Agent("openai:gpt-4o")

@agent.tool_plain
@ExaConnector.agent_tool(
    framework="pydantic_ai",
    inspect_tool="exa_inspect",
    docs_tool="exa_read_docs",
)
async def exa_execute(entity: str, action: str, params: dict | None = None):
    return await connector.execute(entity, action, params or {})

@agent.tool_plain
@ExaConnector.agent_tool(framework="pydantic_ai")
async def exa_inspect():
    return await connector.inspect_connector()

@agent.tool_plain
@ExaConnector.agent_tool(framework="pydantic_ai")
async def exa_read_docs(section: str | None = None):
    return await connector.read_skill_docs(section)

Use the same three-function pattern with framework="langchain", "openai_agents", or "mcp" and that framework's registration decorator. Each value translates connector failures into the framework's own signal:

framework=Tool failures surface as
"pydantic_ai"pydantic_ai.ModelRetry
"langchain"langchain_core.tools.ToolException (set handle_tool_error=True to feed it back to the model)
"openai_agents"the failure message returned to the model as the tool result
"mcp"fastmcp.exceptions.ToolError
"none" (default)airbyte_agent_sdk.AirbyteToolError

On a framework the SDK does not support natively — or in a raw LLM dispatch loop — omit framework= and handle AirbyteToolError yourself:

No framework
from airbyte_agent_sdk import AirbyteToolError
from airbyte_agent_sdk.connectors.exa import ExaConnector
from airbyte_agent_sdk.connectors.exa.models import ExaAuthConfig

connector = ExaConnector(
    auth_config=ExaAuthConfig(
        api_key="<Your Exa API key from dashboard.exa.ai/api-keys>"
    )
)

@ExaConnector.agent_tool(
    inspect_tool="exa_inspect",
    docs_tool="exa_read_docs",
)
async def exa_execute(entity: str, action: str, params: dict | None = None):
    return await connector.execute(entity, action, params or {})

@ExaConnector.agent_tool()
async def exa_inspect():
    return await connector.inspect_connector()

@ExaConnector.agent_tool()
async def exa_read_docs(section: str | None = None):
    return await connector.read_skill_docs(section)

# Advertise all three to the model, using each function's docstring as its description.
handlers = {
    fn.__name__: fn
    for fn in (exa_inspect, exa_read_docs, exa_execute)
}

# `tool_name` and `tool_args` come from the model's tool call in your dispatch loop.
try:
    tool_result = await handlers[tool_name](**tool_args)
except AirbyteToolError as err:
    tool_result = str(err)  # hand the message back to the model as an errored tool result

Each function's docstring carries the guidance the model needs, so pass it through as the tool description wherever you register it.

Legacy alternatives​

These examples are kept for existing integrations. The deprecated ExaConnector.tool_utils pattern loads the connector's full generated catalog into one broad execute tool description instead of letting the agent read skill docs on demand. For new code, use build_connector_tools or ExaConnector.agent_tool above.

Pydantic AI
from pydantic_ai import Agent
from airbyte_agent_sdk.connectors.exa import ExaConnector
from airbyte_agent_sdk.connectors.exa.models import ExaAuthConfig

connector = ExaConnector(
    auth_config=ExaAuthConfig(
        api_key="<Your Exa API key from dashboard.exa.ai/api-keys>"
    )
)

agent = Agent("openai:gpt-4o")

@agent.tool_plain
@ExaConnector.tool_utils
async def exa_execute(entity: str, action: str, params: dict | None = None):
    return await connector.execute(entity, action, params or {})

Authentication​

For all authentication options, see the connector's authentication documentation.

IP allow list​

If your organization restricts access to specific IPs, add the Airbyte Agents IP addresses to your allow list.

Version information​

Connector version: 1.0.0

Was this page helpful?