Skip to main content

Github

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

GitHub is a platform for version control and collaborative software development using Git. This connector provides access to repositories, branches, commits, issues, pull requests, reviews, comments, releases, discussions, organizations, teams, and users for development workflow analysis and project management insights.

Example prompts​

The Github connector is optimized to handle prompts like these.

  • Show me my GitHub profile
  • List my repositories
  • Show me the details of my most recently updated repository
  • List the repositories in my first organization
  • Search for repositories related to machine learning
  • List all my organizations
  • Show me all open issues in my repositories this month
  • Show me the details of the most recently created issue in my first repository
  • Search for issues labeled as bugs in my first repository
  • Find all pull requests created in the past two weeks
  • Show me the details of the most recent pull request in my first repository
  • Search for pull requests that are still open in my first repository
  • List the comments on the most recent pull request in my first repository
  • List the reviews on the most recent pull request in my first repository
  • List the comments on the most recent issue in my first repository
  • Show me the details of the most recent comment on the first issue in my repository
  • List the most recent commits in my first repository and who authored them
  • Show me the details of the most recent commit in my first repository
  • List the branches in my main repository
  • Show me the details of the default branch in my first repository
  • Show me the content of the README file in my first repository
  • Get details about the most recent releases in my organization
  • Show me the details of the latest release in my first repository
  • List all milestones for our current development sprint
  • Show me the details of the first milestone in my repository
  • List all labels in my first repository
  • Show me the details of the first label in my repository
  • List all tags in my first repository
  • Show me the details of the most recent tag in my repository
  • List the stargazers of my first repository
  • Search for users named after my organization
  • List all projects in my first organization
  • Show me the details of the first project
  • List the items in my first project
  • List all unanswered discussions in a repository
  • Show me recent discussions in the General category
  • Search for discussions about releases in my first repository
  • Show me the details of the most recent discussion in my repository
  • List the files and folders in the docs directory of my first repository
  • Show me the details of the first comment on the most recent pull request in my first repository
  • List the public repositories of the user girarda
  • Show me the profile of the GitHub user airbyteio
  • Create a new issue titled 'Fix login bug' in my repository
  • File a new bug report issue in our project repository
  • Open a new feature request issue in the repository
  • Close the most recent issue in my repository as completed
  • Add a comment to the most recent issue saying 'This has been fixed in the latest release'
  • Create a pull request from my latest feature branch to main in my repository

Unsupported prompts​

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

  • Delete an old branch from the repository
  • Schedule a team review for this code
  • Merge a pull request
  • Delete an issue or comment

Entities and actions​

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

EntityActions
RepositoriesGet, List, Search, Context Store Search, Context Store SQL Query
Org RepositoriesList, Context Store Search, Context Store SQL Query
BranchesList, Get, Context Store Search, Context Store SQL Query
CommitsList, Get, Context Store Search, Context Store SQL Query
ReleasesList, Get, Context Store Search, Context Store SQL Query
IssuesList, Get, Search, Create, Update, Context Store Search, Context Store SQL Query, Semantic Search
CommentsCreate, List, Get, Context Store Search, Context Store SQL Query, Semantic Search
Pull RequestsCreate, List, Get, Search, Context Store Search, Context Store SQL Query, Semantic Search
ReviewsList, Context Store Search, Context Store SQL Query
Pr CommentsList, Get, Context Store Search, Context Store SQL Query
LabelsList, Get, Context Store Search, Context Store SQL Query
MilestonesList, Get, Context Store Search, Context Store SQL Query
OrganizationsGet, List, Context Store Search, Context Store SQL Query
UsersGet, List, Search, Context Store Search, Context Store SQL Query
TeamsList, Get, Context Store Search, Context Store SQL Query
TagsList, Get, Context Store Search, Context Store SQL Query
StargazersList, Context Store Search, Context Store SQL Query
ViewerGet, Context Store Search, Context Store SQL Query
Viewer RepositoriesList, Context Store Search, Context Store SQL Query
ProjectsList, Get, Context Store Search, Context Store SQL Query
Project ItemsList, Context Store Search, Context Store SQL Query
DiscussionsList, Get, Search, Context Store Search, Context Store SQL Query
File ContentGet, Context Store Search, Context Store SQL Query
Directory ContentList, Context Store Search, Context Store SQL Query

Github API docs​

See the official Github API reference.

Interfaces​

Use the Github 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": "github"
}'

Describe the connector to see its supported entities and actions:

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

Execute an action:

airbyte-agent connectors execute --json '{
  "workspace": "<your_workspace_name>",
  "name": "github",
  "entity": "repositories",
  "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 GithubConnector 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.github import GithubConnector

connector = connect("github", 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 GithubConnector.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.github import GithubConnector

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

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

@agent.tool_plain
@GithubConnector.agent_tool(
    framework="pydantic_ai",
    inspect_tool="github_inspect",
    docs_tool="github_read_docs",
)
async def github_execute(entity: str, action: str, params: dict | None = None):
    return await connector.execute(entity, action, params or {})

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

@agent.tool_plain
@GithubConnector.agent_tool(framework="pydantic_ai")
async def github_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.github import GithubConnector

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

@GithubConnector.agent_tool(
    inspect_tool="github_inspect",
    docs_tool="github_read_docs",
)
async def github_execute(entity: str, action: str, params: dict | None = None):
    return await connector.execute(entity, action, params or {})

@GithubConnector.agent_tool()
async def github_inspect():
    return await connector.inspect_connector()

@GithubConnector.agent_tool()
async def github_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 (github_inspect, github_read_docs, github_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 GithubConnector.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 GithubConnector.agent_tool above.

Pydantic AI
from pydantic_ai import Agent
from airbyte_agent_sdk import connect
from airbyte_agent_sdk.connectors.github import GithubConnector

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

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

@agent.tool_plain
@GithubConnector.tool_utils
async def github_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.github import GithubConnector
from airbyte_agent_sdk.types import AirbyteAuthConfig

connector = GithubConnector(
    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.github import GithubConnector
from airbyte_agent_sdk.connectors.github.models import GithubPersonalAccessTokenAuthConfig

connector = GithubConnector(
    auth_config=GithubPersonalAccessTokenAuthConfig(
        token="<GitHub personal access token (fine-grained or classic)>"
    )
)

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 GithubConnector.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.github import GithubConnector
from airbyte_agent_sdk.connectors.github.models import GithubPersonalAccessTokenAuthConfig

connector = GithubConnector(
    auth_config=GithubPersonalAccessTokenAuthConfig(
        token="<GitHub personal access token (fine-grained or classic)>"
    )
)

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

@agent.tool_plain
@GithubConnector.agent_tool(
    framework="pydantic_ai",
    inspect_tool="github_inspect",
    docs_tool="github_read_docs",
)
async def github_execute(entity: str, action: str, params: dict | None = None):
    return await connector.execute(entity, action, params or {})

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

@agent.tool_plain
@GithubConnector.agent_tool(framework="pydantic_ai")
async def github_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.github import GithubConnector
from airbyte_agent_sdk.connectors.github.models import GithubPersonalAccessTokenAuthConfig

connector = GithubConnector(
    auth_config=GithubPersonalAccessTokenAuthConfig(
        token="<GitHub personal access token (fine-grained or classic)>"
    )
)

@GithubConnector.agent_tool(
    inspect_tool="github_inspect",
    docs_tool="github_read_docs",
)
async def github_execute(entity: str, action: str, params: dict | None = None):
    return await connector.execute(entity, action, params or {})

@GithubConnector.agent_tool()
async def github_inspect():
    return await connector.inspect_connector()

@GithubConnector.agent_tool()
async def github_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 (github_inspect, github_read_docs, github_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 GithubConnector.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 GithubConnector.agent_tool above.

Pydantic AI
from pydantic_ai import Agent
from airbyte_agent_sdk.connectors.github import GithubConnector
from airbyte_agent_sdk.connectors.github.models import GithubPersonalAccessTokenAuthConfig

connector = GithubConnector(
    auth_config=GithubPersonalAccessTokenAuthConfig(
        token="<GitHub personal access token (fine-grained or classic)>"
    )
)

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

@agent.tool_plain
@GithubConnector.tool_utils
async def github_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: 0.1.19

Was this page helpful?