When you ask an AI agent to “send that email in Gmail” or “post this to Slack,” authentication is the hardest part. Whose permissions does it act with, and as which account? Where are the tokens stored, and how do you record who did what? Arcade is an MCP (Model Context Protocol) runtime that takes on all three problems at once ─ authorization, execution, and governance ─ for AI agents. Built by Arcade AI in the United States, it can be tried on a free plan. Agents authenticate through OAuth as the real end user and operate strictly within the permissions that person already has, while administrators can trace every action through a central audit log.
Key Features
- OAuth that acts as the real user: Agents authenticate as the person they are acting for, not through a shared admin token. Access tokens, API keys, and secrets stay isolated inside the Arcade runtime and are never handed to the agent’s context
- 7,500+ pre-built tools: A large catalog spanning 81 MCP servers, covering the mainstays of business systems ─ Google Workspace, Microsoft 365, Slack, Notion, Jira, Linear, Salesforce, HubSpot, Stripe, GitHub, Postgres, Snowflake, and more
- Works with major frameworks and MCP clients: SDKs for Python and TypeScript, integrations with agent frameworks such as LangChain, CrewAI, OpenAI Agents, Google ADK, Mastra, and the Vercel AI SDK, plus MCP clients including Claude Desktop, Claude Code, Cursor, and VS Code
- Central governance through audit logs: The control plane consolidates “which agent did what, on behalf of whom, in which system,” with OpenTelemetry-format audit logs, a shared registry, version control, and tool visibility filtering
- Flexible deployment: Beyond the fully managed Arcade Cloud, it can run in a VPC, on-premises, or in air-gapped environments, with distribution via the AWS/Azure marketplaces or Helm-based self-hosting
- Enterprise-grade security: SOC 2 compliant, with SSO, RBAC, and audit logs, and connections to existing identity providers such as Auth0, Okta, Microsoft Entra ID, Clerk, and Stytch
- Bring your own tools: In addition to the prepared catalog, you can register custom tools or existing in-house MCP servers under Arcade’s governance
Pricing
| Plan | Monthly price | What’s included |
|---|---|---|
| Free | $0 | 2,000 auth events/month and 2,000 tool calls/month. Fully managed on Arcade Cloud, community support (Discord and GitHub). No credit card required |
| Team | $25 + usage | A $25/month platform fee plus $0.10 per auth event and $0.01 per tool call. Next-business-day email support |
| Enterprise | Contact sales | Annual bundles of auth events and tool calls, discounted rates at scale, flexible deployment across cloud/VPC/air-gapped environments, 24/7 SLA-backed support, a dedicated team, SSO, RBAC, audit logs, and private registry access |
Pricing is current as of August 2026. Check the official site for the latest information.
Pros & Cons
✅ Pros
- Removes the need to build OAuth plumbing yourself, freeing you from maintaining authentication code
- Runs within each user’s own permissions instead of requiring an all-powerful “agent account,” keeping your permission model consistent with existing access control
- A large catalog of pre-built tools means you rarely have to write business-system integrations from scratch
- Audit logs and RBAC are built in from the start, which makes internal approval and security review easier to clear
- Deployable to VPC, on-premises, or air-gapped environments, so organizations that cannot send data outside can still consider it
- The free plan offers 2,000 auth events and 2,000 tool calls per month to experiment with
⚠️ Cons
- Usage-based pricing makes costs hard to forecast while your agent’s execution volume is still unpredictable
- Assumes familiarity with MCP and OAuth; this is not a no-code tool
- SDKs cover only Python and TypeScript, so other languages have to go through the HTTP API
- Concentrating your authentication layer in one vendor means exposure to its availability and specification changes
- Purpose-built for having agents carry out real work, which makes it somewhat heavy for plain API automation
Comparison with Similar Services
| Criteria | Arcade | Composio | Zapier MCP | Pipedream Connect |
|---|---|---|---|---|
| Positioning | Authorization and execution runtime for agents | Tool integration platform for agents | Opens an existing automation platform through MCP | Auth and integration platform for developers |
| Auth approach | OAuth as the real end user | Per-user connections | Connections tied to the Zapier account | Per-end-user connections |
| Deployment | Cloud / VPC / on-premises / air-gapped | Primarily cloud | Cloud only | Mainly cloud |
| Governance | Audit logs, RBAC, and SSO as standard | Available (plan dependent) | Follows Zapier’s admin features | Available |
| Best for | Agent development under governance requirements | Developers wiring up many tools quickly | Teams leveraging existing Zapier assets | Developers embedding integrations in their own product |
Who Is It For
- Developers building internal AI agents who are losing time to OAuth and token management
- Teams that cannot get internal approval until they can explain whose permissions an agent acted with
- Engineers who have built agents on LangChain or OpenAI Agents and want to carve out only the external tool integration
- Companies that cannot send data to an external cloud and must run their agent platform in a VPC or on-premises
- Anyone who wants to operate business systems safely from an MCP client such as Claude Desktop or Cursor
Summary
Arcade bundles into a single runtime the three problems you always hit once AI agents start touching real business systems: authorization, execution, and governance. Its design of authenticating through OAuth as the actual end user, combined with audit logs and RBAC out of the box, makes it far easier to move agents past the prototype stage and into an organization’s control requirements. A good starting point is the free plan’s monthly allowance of 2,000 calls, used to confirm that the tools your use case needs are already there.