arcade-mcp

Tools can declare:

1. Which OAuth scopes they need, and Arcade handles the rest: prompting the end user to authorize, storing and refreshing tokens, scoping them per call
2. API keys or any other secrets, and Arcade hosts the secrets securely in an encrypted environment.

The secrets (OAuth tokens, API keys, etc) are securely injected by Arcade into your tool call at runtime, so that your tool can authorize requests to upstream APIs, for example. The client and the LLM never see the secret values.

A tool that reads the user's GitHub repos is one decorator away:

from arcade_mcp_server import MCPApp, Context
from arcade_mcp_server.auth import GitHub

app = MCPApp(name="gh", version="1.0.0")

@app.tool(requires_auth=GitHub(scopes=["repo"]))
async def list_my_repos(context: Context) -> list[str]:
    """List the authenticated user's GitHub repositories."""
    token = context.get_auth_token_or_empty()
    ...

When the tool is invoked through Arcade Cloud, the user is presented with a URL to complete the OAuth challenge in their browser. On success, the token is injected into context for that call. Subsequent calls reuse and refresh the token automatically.

22 helper classes ship with the framework for popular providers:

Asana, Atlassian, Attio, ClickUp, Discord, Dropbox, Figma, GitHub, Google, Hubspot, Linear, LinkedIn, Microsoft, , , , , , , , .

arcade-mcp is the Python framework for building Model Context Protocol servers and the tools that run inside them. It powers the 7,500+ prebuilt tools across 81 MCP servers at Arcade.dev, and is open-sourced so you can build your own.

Use it when you need MCP tools that aren't already in the prebuilt catalog: internal APIs, custom OAuth providers, domain-specific integrations.

Highlights:

• Decorator API covering the full MCP spec: tools, resources, prompts, sampling, elicitation, progress, logging.
• Authorized tool calling. Declare requires_auth=GitHub(scopes=["repo"]) and Arcade handles OAuth, token refresh, and per-call scoping. The client and the LLM never see the token.
• Vendor-neutral. Any MCP client, any LLM, any agent framework (LangChain, Mastra, Pydantic AI, CrewAI, Google ADK, OpenAI Agents).
• arcade evals for testing tool-call accuracy against real LLMs.
• arcade deploy for one-command hosting on Arcade Cloud.