Arcade Documentation — Authorized Tool Calling and MCP Servers for AI Agents
by Arcade
The OAuth and per-action authorization layer most agent frameworks leave you to write yourself.
Overview
Arcade documents the layer almost every agent tutorial skips: how an agent acquires and holds a real user's credentials without becoming a security incident. The docs are organised into five sections. Get Started covers setup, quickstarts, agent framework integration and MCP client configuration. Build covers tool calling, creating your own tools, user-facing agents and MCP servers. Operate covers deployment options, identity management and governance. Resources holds the integration catalogue, examples, a glossary and FAQ, and References covers the API, CLI, auth providers and telemetry. The platform's three claims are authorization — managing OAuth flows, user tokens, API keys and secrets, enforcing identity-provider policy and per-action authorization at runtime; execution — roughly 7,500 agent-optimized tools across 81 MCP servers spanning Google Workspace, Microsoft 365, Notion, Slack, Teams, Salesforce, HubSpot, Postgres, Snowflake and MongoDB; and governance — a shared registry with version control, visibility filtering and OpenTelemetry audit logs. Documented integrations exist for LangChain, OpenAI Agents, CrewAI, AG2, Vercel AI SDK, Google ADK, Mastra and TanStack AI, in Python and TypeScript. The underlying `arcade-mcp` Python framework is MIT-licensed and self-hostable, with a decorator-based API covering the MCP specification, 22 pre-built provider helper classes and a scaffolding CLI. The hosted free tier allows 2,000 auth events and 2,000 tool calls per month.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Documentation
- Cost
- Freemium
- Duration
- Reference docs — quickstart in ~30 minutes, ~3 hours for the Build and Operate guides
- Provider
- Arcade
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Run an OAuth flow so an agent acts as the end user, not a shared account
- ✓Scope authorization per action so one tool cannot escalate another's privileges
- ✓Build a custom MCP server in Python using a decorator-based tool API
- ✓Wire authorized tools into LangChain, CrewAI, Google ADK or the Vercel AI SDK
- ✓Add human-in-the-loop consent before an agent performs a real-world action
- ✓Deploy tool servers to managed cloud, a VPC or an air-gapped environment
- ✓Audit agent actions via OpenTelemetry logs and a shared, version-controlled registry
- ✓Keep API keys and secrets out of prompts, agent code and conversation history
Highlights
- •Solves the problem most agent frameworks skip: acting as the user rather than as a shared service account
- •Roughly 7,500 pre-built tools across 81 MCP servers — Google Workspace, Slack, Notion, Salesforce, Postgres and more
- •Framework-agnostic, with documented integrations for LangChain, OpenAI Agents, CrewAI, AG2, Google ADK, Mastra and Vercel AI SDK
- •The arcade-mcp Python framework is MIT-licensed and self-hostable; the managed runtime is optional, not required
- •Free tier of 2,000 auth events and 2,000 tool calls per month with no card required, so the docs are testable end to end
Who It’s For
Best For
- ✓Engineers whose agent needs to read a specific user's Gmail, Slack or Salesforce
- ✓Teams that failed a security review over how their agent handles credentials
- ✓Developers building MCP servers who do not want to write OAuth plumbing
- ✓Platform teams that need audit logs and a governed tool registry across many agents
Prerequisites
- •Python 3.10+ or TypeScript, and an agent or LLM application already running
- •Understanding of tool/function calling and Model Context Protocol basics
- •Familiarity with OAuth concepts such as scopes, consent and refresh tokens
FAQ
What is Arcade Documentation — Authorized Tool Calling and MCP Servers for AI Agents?
Official documentation for Arcade, an actions runtime that lets an agent take real actions in a user's own accounts rather than through a shared service credential. It covers OAuth flows, per-action authorization, secret handling, building MCP servers in Python, and governance via audit logs and a shared registry. Written for engineers whose agent needs to touch Gmail, Slack, Salesforce or a production database safely.
Is Arcade Documentation — Authorized Tool Calling and MCP Servers for AI Agents free?
Arcade Documentation — Authorized Tool Calling and MCP Servers for AI Agents offers free content, with paid options for certificates or premium features.
What level is Arcade Documentation — Authorized Tool Calling and MCP Servers for AI Agents for?
Arcade Documentation — Authorized Tool Calling and MCP Servers for AI Agents is aimed at a intermediate audience. Recommended background: Python 3.10+ or TypeScript, and an agent or LLM application already running, Understanding of tool/function calling and Model Context Protocol basics, Familiarity with OAuth concepts such as scopes, consent and refresh tokens.
How long does Arcade Documentation — Authorized Tool Calling and MCP Servers for AI Agents take?
Expect roughly Reference docs — quickstart in ~30 minutes, ~3 hours for the Build and Operate guides. Most learners work through it at their own pace.
What will I learn from Arcade Documentation — Authorized Tool Calling and MCP Servers for AI Agents?
You'll learn: Run an OAuth flow so an agent acts as the end user, not a shared account; Scope authorization per action so one tool cannot escalate another's privileges; Build a custom MCP server in Python using a decorator-based tool API; Wire authorized tools into LangChain, CrewAI, Google ADK or the Vercel AI SDK; Add human-in-the-loop consent before an agent performs a real-world action; Deploy tool servers to managed cloud, a VPC or an air-gapped environment; Audit agent actions via OpenTelemetry logs and a shared, version-controlled registry; Keep API keys and secrets out of prompts, agent code and conversation history.
Topics
Sources
This page was written from 3 sources, 2 on domains other than docs.arcade.dev.