One

What One MCP server does

One MCP server gives an AI agent a single interface for working with connected applications. The service supports more than 700 platforms, including examples such as Gmail, Slack, Shopify, HubSpot, Stripe, Linear, and QuickBooks. It can identify the operation needed for a task, return the relevant action documentation, and execute the API request on a connected platform.

The server keeps its MCP surface small: three tools are available regardless of how many platforms or actions are connected. Access can be narrowed by connection, action, and permission level. Connections report their available access so an agent can determine what it is allowed to do before attempting an operation.

How it works

Agents first inspect available integrations and active connections with list_one_integrations. They can then use find_one_actions to describe one or more operations. The response selects an action, includes its documentation, identifies a substitute when selection is uncertain, and may list alternatives. Large documentation responses can be expanded by requesting a section, the full document, or its table of contents.

execute_one_action runs an approved API action against a connected platform. Requests are proxied through One, so platform API keys and OAuth tokens do not need to be individually managed by the agent. The service also redacts secrets from requests.

One MCP server supports both local and remote deployment. The local package runs through the One CLI or the @withone/mcp npm package. The hosted endpoint at https://mcp.withone.ai/mcp uses Streamable HTTP and authenticates through One's OAuth flow.

Setup and configuration

For local use, install the CLI globally, run one init, and provide an API key from the One dashboard when prompted. Manual setup installs @withone/mcp and sets ONE_SECRET. The CLI can configure Claude Desktop, Cursor, Claude Code, Windsurf, or another MCP-compatible agent.

The remote configuration only needs the hosted URL. On first connection, the client opens a browser for authentication and authorization. This avoids local npm installation and does not require ONE_SECRET.

Optional identity settings restrict returned connections to a user, team, organization, or project. ONE_PERMISSIONS limits actions to read, write, or admin access. ONE_CONNECTION_KEYS and ONE_ACTION_IDS restrict visible and executable connections or actions. Setting ONE_KNOWLEDGE_AGENT=true removes execution and makes the server documentation-only.

Tools and capabilities

• list_one_integrations: Shows platforms, active connections, and the access level granted to each connection.
• find_one_actions: Finds actions for described operations and returns their documentation, alternatives, or additional requested sections.
• execute_one_action: Executes an API action on a connected platform.
• Fine-grained filtering by permission, connection key, and action ID.
• Knowledge-only operation for agents that should generate integration code without executing actions.

Limitations and notes

The older search_one_platform_actions and get_one_action_knowledge tools are removed in the current version; find_one_actions replaces them. Execution depends on having an appropriate active connection and sufficient configured access. Restricting permissions or allowlists can cause actions to be omitted or replaced with another permitted candidate. In knowledge-only mode, the server does not expose execute_one_action.

The README documents local setup and the hosted remote endpoint, but does not specify a license. The examples demonstrate natural-language tasks and generated application code; the actual available operations depend on the connected platforms and One configuration.