Source: Arcade authorization docs

Arcade's docs on authorized tool calling: per-user OAuth via tools.authorize, server-level vs tool-level auth for MCP servers, and user verification against auth-link phishing.

Contents

Summary

Arcade is a tool runtime and credential broker. Your app identifies the end user with a user_id; tools.authorize(tool_name, user_id) returns either a completed status or a URL the user must visit; after consent, tools.execute(tool_name, user_id) runs the tool with the stored token injected server-side. For MCP servers it separates front-door (resource-server) auth from per-tool third-party OAuth, and in production it requires your app to confirm that the person finishing an OAuth flow is the user who started it.

Key claims

  • Flow: authorize → if status is not completed, show url as-is → wait_for_completion → execute with only tool name and user_id. Authorization is remembered per user until it expires or is revoked; it can be done ahead of time.
  • Server-level auth: the MCP server is an OAuth 2.1 protected resource; middleware validates the Bearer token on every request and takes user_id from sub. Not needed for stdio.
  • Tool-level auth: declared per tool, e.g. GitHub with scopes repo, read:user, so each tool requests only its scopes; the token reaches tool code through the Context object, not through tool parameters.
  • User verification: after a first authorization Arcade redirects to your verifier route with a flow_id; your server reads the signed-in user from its own session (never request input) and calls confirmUser; a mismatch returns user_mismatch. This defends against phishing where a victim completes a flow started under an attacker’s ID.
  • Production: use a custom verifier and your own OAuth app credentials; Arcade’s default OAuth apps are for development and single-tenant use. Call the verification API only from the server.

Feeds

Suggest changes →