Choose how to use Airbyte Agents

FreeIndividualTeamCustom Compare plans

Airbyte Agents offers several interfaces. They all share the same platform, the same connectors, and the same Context Store, so you can start with one and add others as your needs grow.

Use the flowchart below to find the best starting point, then read the section that matches your path.

At a glance

	Web app	MCP server	SDK	CLI	API
Install required	No	No	uv add airbyte-agent-sdk	Binary download or brew	No
Language / platform	Browser	Any MCP-capable client	Python	Any shell	Any HTTP client
Auth model	Browser login	OAuth (browser popup)	Client ID + secret	Browser login or --manual	Client ID + secret
Best for	No-code exploration, automations	Conversational AI agents (Claude, ChatGPT, Cursor)	Python-native agent frameworks	Shell scripts, CI, agent harnesses	Non-Python backends, custom admin
Programmatic control	No	Agent-driven	Full	Full	Full
Context Store access	Yes	Yes	Yes	Yes	Yes
Credential handling	Browser widget	Browser widget (via agent)	Environment variables	Browser widget + settings file	HTTP headers

MCP server

Best for: Users of Claude, Cursor, ChatGPT, VS Code, or any agent that supports the Model Context Protocol.

The MCP server is a remote, Airbyte-hosted server that gives MCP-capable agents authenticated access to your connected data. You have nothing to install. Add the server URL to your agent's MCP configuration, authenticate with your Airbyte account, and your agent can immediately read and write data across every connector in your workspace.

Choose MCP when you are using an off-the-shelf conversational client that supports MCP and you want zero-install access to your data. For coding agents and harnesses that can run shell commands, see CLI.

Get started: see the MCP server docs for setup instructions for Claude Code, Cursor, VS Code, Claude Desktop, ChatGPT, and other clients.

Python SDK

Best for: Python developers building custom agents with frameworks like Pydantic AI, LangChain, or FastMCP.

The Python SDK (airbyte-agent-sdk) gives you typed connectors, automatic credential handling, and patterns for exposing connectors as tools to any AI agent framework. Install the SDK, authenticate with your Airbyte API credentials, and start executing operations in your own code.

Choose the SDK when you're writing Python and want typed, in-process access to connectors. If your stack isn't Python, use the API. If you prefer a shell binary over a library import, see the CLI.

uv add airbyte-agent-sdk

Get started: follow one of the step-by-step tutorials in the Developer Quickstart:

• Pydantic AI tutorial

• LangChain tutorial

• FastMCP tutorial

Web app

Best for: Non-developers, operations teams, and anyone who wants to explore Airbyte Agents without writing code.

The web app at app.airbyte.ai is the fastest way to get started. Describe what you need in natural language, and an Airbyte-hosted agent picks the right connectors, makes the necessary tool calls, and replies with an answer grounded in your data.

Use Chats to have interactive conversations with an agent. Ask a question, iterate on the answer, and explore your data in real time.

Get started: Sign up at app.airbyte.ai, add a connector on the Connectors page, and open New Chat.

CLI

Best for: Shell scripts, CI jobs, and agent harnesses that can run command-line tools.

The CLI (airbyte-agent) gives you a terminal interface for listing workspaces, adding connectors through browser credential flows, describing connector schemas, and executing connector actions. It accepts JSON input, prints JSON output, and includes schema discovery for agent workflows.

Choose the CLI when you are using a coding agent or harness that can run shell commands. The CLI runs locally, authenticates via a browser flow or --manual mode, and returns structured JSON for programmatic consumption. It is especially useful when output is large, work requires multiple steps, or an operation may run for a long time.

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

Get started: see the CLI docs for install, authentication, connector creation, execution, and agent usage.

Agent API

Best for: backend engineers, non-Python stacks, custom admin flows, and embedding the authentication module in your application.

The Agent API exposes REST endpoints for managing connectors, tokens, and executing operations from any language or backend service. Use it when you need programmatic control over Airbyte Agents from a stack that isn't Python, or when you're building a custom integration layer.

Choose the API when you need HTTP-level control, are working outside Python, or are building admin tooling (workspace management, credential embedding). If you're in Python, the SDK wraps the same endpoints with a typed interface.

Get started: see the API docs for authentication, connector management, and execution endpoints. The Developer Quickstart also covers common patterns.

MCP server vs. CLI for AI agents

Both the MCP server and the CLI connect AI agents to your data, but they differ in how the agent communicates with Airbyte:

• MCP server — The agent talks to an Airbyte-hosted server over the Model Context Protocol. Nothing is installed locally. Authentication uses OAuth through a browser popup.
• CLI — The agent shells out to a local airbyte-agent binary. Commands accept JSON input and print JSON output. Install the binary with the install script, then authenticate with a browser login or --manual for headless machines.

The CLI gives coding agents and harnesses several practical advantages:

• The install script includes an installable agent skill with fuller, more prescriptive guidance than MCP tool descriptions and server instructions can provide. See Use the CLI with AI agents.
• CLI output composes with shell tools such as grep, jq, and tail, as well as other command-line programs, so the agent can shape output for each step.
• The CLI does not truncate command output, while some MCP clients truncate tool-call results.
• The agent can run as many shell commands as the task needs in one turn, while some MCP clients cap tool calls per turn or require approval between calls.
• Long-running work can finish in a shell command, while MCP tool calls are subject to client request timeouts.

Use MCP for an off-the-shelf conversational client when zero installation matters. Use the CLI for coding agents and harnesses that can run shell commands, especially when output is large, work is multi-step, or tasks are long-running.

All paths lead to the same data

Whichever interface you choose, your agents work with the same connectors, the same credentials, and the same Context Store. A connector you add in one interface is immediately available through the others. You can mix and match interfaces as your needs evolve.

For a deeper look at how the platform is organized, see System architecture.