Skip to main content

Gmail

The Gmail agent connector is a Python package that equips AI agents to interact with Gmail 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.

Gmail is Google's email service that provides email sending, receiving, and organization capabilities. This connector provides access to messages, threads, labels, drafts, and user profile information. It supports read operations for listing and retrieving email data, as well as write operations including sending messages, managing drafts, modifying message labels, and creating or updating labels.

Example prompts​

The Gmail connector is optimized to handle prompts like these.

  • List my recent emails
  • Show me unread messages in my inbox
  • Get the details of a specific email
  • List all my Gmail labels
  • Show me details for a specific label
  • List my email drafts
  • Get the content of a specific draft
  • List my email threads
  • Show me the full thread for a conversation
  • Get my Gmail profile information
  • Send an email to someone
  • Create a new email draft
  • Archive a message by removing the INBOX label
  • Mark a message as read
  • Mark a message as unread
  • Move a message to trash
  • Create a new label
  • Update a label name or settings
  • Delete a label
  • Search for messages matching a query
  • Find emails from a specific sender
  • Show me emails with attachments
  • Show me my latest emails
  • Find recent unread emails about {topic}
  • Show the conversation with {person} about {topic}

Unsupported prompts​

The Gmail connector isn't currently able to handle prompts like these.

  • Attach a file to an email
  • Forward an email to someone
  • Create a filter or rule
  • Manage Gmail settings
  • Access Google Calendar events
  • Manage contacts

Entities and actions​

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

EntityActions
ProfileGet, Context Store Search, Context Store SQL Query
MessagesList, Get, Create, Update, Context Store Search, Context Store SQL Query
LabelsList, Create, Get, Update, Delete, Context Store Search, Context Store SQL Query
DraftsList, Create, Get, Update, Delete, Context Store Search, Context Store SQL Query
Drafts SendCreate
ThreadsList, Get, Context Store Search, Context Store SQL Query
Messages TrashCreate
Messages UntrashCreate

Gmail API docs​

See the official Gmail API reference.

Interfaces​

Use the Gmail 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": "gmail"
}'

Describe the connector to see its supported entities and actions:

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

Execute an action:

airbyte-agent connectors execute --json '{
  "workspace": "<your_workspace_name>",
  "name": "gmail",
  "entity": "profile",
  "action": "get"
}'

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 GmailConnector 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.gmail import GmailConnector

connector = connect("gmail", 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 GmailConnector.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.gmail import GmailConnector

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

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

@agent.tool_plain
@GmailConnector.agent_tool(
    framework="pydantic_ai",
    inspect_tool="gmail_inspect",
    docs_tool="gmail_read_docs",
)
async def gmail_execute(entity: str, action: str, params: dict | None = None):
    return await connector.execute(entity, action, params or {})

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

@agent.tool_plain
@GmailConnector.agent_tool(framework="pydantic_ai")
async def gmail_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.gmail import GmailConnector

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

@GmailConnector.agent_tool(
    inspect_tool="gmail_inspect",
    docs_tool="gmail_read_docs",
)
async def gmail_execute(entity: str, action: str, params: dict | None = None):
    return await connector.execute(entity, action, params or {})

@GmailConnector.agent_tool()
async def gmail_inspect():
    return await connector.inspect_connector()

@GmailConnector.agent_tool()
async def gmail_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 (gmail_inspect, gmail_read_docs, gmail_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 GmailConnector.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 GmailConnector.agent_tool above.

Pydantic AI
from pydantic_ai import Agent
from airbyte_agent_sdk import connect
from airbyte_agent_sdk.connectors.gmail import GmailConnector

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

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

@agent.tool_plain
@GmailConnector.tool_utils
async def gmail_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.gmail import GmailConnector
from airbyte_agent_sdk.types import AirbyteAuthConfig

connector = GmailConnector(
    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.gmail import GmailConnector
from airbyte_agent_sdk.connectors.gmail.models import GmailAuthConfig

connector = GmailConnector(
    auth_config=GmailAuthConfig(
        access_token="<Your Google OAuth2 Access Token (optional, will be obtained via refresh)>",
        refresh_token="<Your Google OAuth2 Refresh Token>",
        client_id="<Your Google OAuth2 Client ID>",
        client_secret="<Your Google OAuth2 Client Secret>"
    )
)

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 GmailConnector.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.gmail import GmailConnector
from airbyte_agent_sdk.connectors.gmail.models import GmailAuthConfig

connector = GmailConnector(
    auth_config=GmailAuthConfig(
        access_token="<Your Google OAuth2 Access Token (optional, will be obtained via refresh)>",
        refresh_token="<Your Google OAuth2 Refresh Token>",
        client_id="<Your Google OAuth2 Client ID>",
        client_secret="<Your Google OAuth2 Client Secret>"
    )
)

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

@agent.tool_plain
@GmailConnector.agent_tool(
    framework="pydantic_ai",
    inspect_tool="gmail_inspect",
    docs_tool="gmail_read_docs",
)
async def gmail_execute(entity: str, action: str, params: dict | None = None):
    return await connector.execute(entity, action, params or {})

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

@agent.tool_plain
@GmailConnector.agent_tool(framework="pydantic_ai")
async def gmail_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.gmail import GmailConnector
from airbyte_agent_sdk.connectors.gmail.models import GmailAuthConfig

connector = GmailConnector(
    auth_config=GmailAuthConfig(
        access_token="<Your Google OAuth2 Access Token (optional, will be obtained via refresh)>",
        refresh_token="<Your Google OAuth2 Refresh Token>",
        client_id="<Your Google OAuth2 Client ID>",
        client_secret="<Your Google OAuth2 Client Secret>"
    )
)

@GmailConnector.agent_tool(
    inspect_tool="gmail_inspect",
    docs_tool="gmail_read_docs",
)
async def gmail_execute(entity: str, action: str, params: dict | None = None):
    return await connector.execute(entity, action, params or {})

@GmailConnector.agent_tool()
async def gmail_inspect():
    return await connector.inspect_connector()

@GmailConnector.agent_tool()
async def gmail_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 (gmail_inspect, gmail_read_docs, gmail_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 GmailConnector.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 GmailConnector.agent_tool above.

Pydantic AI
from pydantic_ai import Agent
from airbyte_agent_sdk.connectors.gmail import GmailConnector
from airbyte_agent_sdk.connectors.gmail.models import GmailAuthConfig

connector = GmailConnector(
    auth_config=GmailAuthConfig(
        access_token="<Your Google OAuth2 Access Token (optional, will be obtained via refresh)>",
        refresh_token="<Your Google OAuth2 Refresh Token>",
        client_id="<Your Google OAuth2 Client ID>",
        client_secret="<Your Google OAuth2 Client Secret>"
    )
)

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

@agent.tool_plain
@GmailConnector.tool_utils
async def gmail_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: 0.2.0

Was this page helpful?