Arcade MCP

What arcade-mcp does

arcade-mcp is an open-source Python framework for developing Model Context Protocol servers and the capabilities they provide to AI agents. It lets developers define tools with decorators and supports the broader MCP feature set, including resources, prompts, sampling, elicitation, progress reporting, and logging.

The framework is intended for custom capabilities that are not already available in a prebuilt catalog. Typical targets include internal APIs, specialized business systems, and services requiring custom authentication. It can also be used to build standalone servers that run directly with MCP clients.

How it works

A server is created with MCPApp, then functions are registered as tools with decorators. Tools can receive a Context object, return typed values, and declare authentication requirements. The application can run over stdio or HTTP+SSE, allowing local client processes and HTTP-capable MCP clients to connect.

Authorized tools declare the scopes or secrets they need. Arcade can prompt users through an OAuth flow, store and refresh tokens, and inject credentials into the tool invocation without exposing secret values to the client or language model. Twenty-three provider helpers cover services including GitHub, Google, Slack, Microsoft, Notion, Linear, Dropbox, Zoom, and others. Generic OAuth2 support is available for providers outside that set.

Servers may run independently with developer-managed credentials or be deployed to Arcade Cloud. Cloud deployment handles required secret discovery and provisioning, while Arcade Engine supports authorized tool-call flows for end users.

Setup and configuration

The project requires Python 3.10 or newer and uv. Install the CLI with:

uv tool install arcade-mcp

Create a starter project using arcade new my_server. The generated project includes a pyproject.toml, an environment-variable example file, and a sample server.py. A minimal application creates an MCPApp, adds a decorated function, and calls app.run(transport="stdio").

Run a local server with uv run server.py. Passing http starts HTTP+SSE mode, with API documentation available at the local server's /docs path. The CLI can configure Claude Desktop, Cursor, and VS Code for the corresponding stdio or HTTP entrypoint.

For hosted operation, authenticate with arcade login and deploy with arcade deploy. Commands for server logs, status, listing, and dashboard access support basic management and observability.

Tools and capabilities

• Decorator-based MCP tools and application setup
• MCP resources, prompts, sampling, elicitation, progress, and logging
• OAuth scope declarations and runtime credential injection
• Provider helpers for 23 OAuth services
• Generic OAuth2 authentication for other APIs
• Local stdio and HTTP+SSE transports
• arcade new project scaffolding
• arcade evals for testing tool-call accuracy with real language models
• arcade deploy for Arcade Cloud hosting
• Resource Server Auth for protecting production HTTP endpoints with OAuth 2.1 bearer tokens

Limitations and notes

The framework is for building servers and tools rather than providing a fixed end-user integration by itself. Developers still need to implement the upstream API calls and tool behavior. Standalone deployments require developers to supply their own access tokens for authenticated tools, while production HTTP deployments should protect endpoints with Resource Server Auth. The documented local client setup covers Claude Desktop, Cursor, and VS Code, with stdio intended for Claude Desktop and command-line tools and HTTP+SSE intended for Cursor and VS Code.