Arcade

Operator's take

The gap between "AI assistant" and "AI agent" is almost always auth. You can prompt your way to a perfect plan, but the moment the agent needs to actually touch a calendar, send an email, or write a record back to a CRM, you're staring at OAuth flows, token refresh logic, and permission scopes. Most teams hand-wave past this until it bites them. Arcade is built specifically to plug that gap: it sits between your AI model and the services it needs to touch, handles the auth layer, and exposes a clean set of tools the model can call.

What this means practically is that you can wire up an AI agent to Google Workspace, Slack, Salesforce, or your own APIs without writing the auth scaffolding yourself. Arcade handles the full auth handshake through your existing identity provider, keeps credentials isolated inside the runtime, and gives you a central control plane where you can see exactly what action ran, on behalf of which user, in which system. For teams building production agents, this collapses weeks of auth plumbing into a configuration problem — and gets you through security review without a custom compliance story.

The honest caveat: Arcade rewards teams who already understand what an API integration means, even if they're not the ones writing the code. If your team thinks in terms of webhooks and data flows, you'll get value quickly. If you're starting from "I want AI to do stuff," the abstraction is real but the mental model still needs to be there. The integration library also matters — if your core systems aren't in their connector catalog, you're back to custom development anyway.

What it's good at

• Agent authorization via your IdP — agents act on behalf of real users through your existing identity provider; credentials never leave the runtime, no shared service accounts.
• MCP runtime — Arcade is the infrastructure layer for MCP tool execution; the team authored the MCP tool authorization spec and sits on MCP security/governance steering committees.
• Pre-built tool catalog — 8,000+ MCP tools across 100+ integrations (Google Workspace, Slack, Salesforce, Microsoft, and more; self-described as the largest agent-ready tool catalog in production); custom tools buildable in Python or JavaScript.
• Central control plane — audit every agent action (what ran, which user, which system) without slowing the teams shipping agents; one place to set policy.
• SOC 2 compliant — SSO, RBAC, and full audit logs included at Enterprise tier; compliance-ready out of the box.
• Flexible deployment — cloud, on-prem, air-gapped, or hybrid; data stays where you control it.
• Usage-based pricing, no seat licenses — free Hobby tier (100 user challenges, 1,000 standard tool executions, 50 pro executions, one Arcade-hosted MCP server); Growth is metered from there (600 user challenges then $0.05 each, 2,000 standard executions then $0.01 each, 100 pro executions then $0.50 each, hosted MCP servers at $0.05/hour); Enterprise moves to custom pricing.

What it's not

• Not a zero-concept tool — teams need to understand the shape of API integrations and agent tool-calling even though they don't have to write the auth code; the mental model is a prerequisite.
• Not a general automation platform — Arcade is the auth/runtime layer for AI agents specifically; if you need broad app-to-app automation without AI in the loop, Zapier or Make are a better fit.
• Not free at compliance-grade — audit logs, RBAC, SSO, and dedicated tenant isolation sit at Enterprise tier; evaluate the pricing shape early if your security review requires them.
• Not useful if your systems aren't in their tool catalog — custom services require custom tool builds; the out-of-the-box value degrades as soon as you leave supported integrations.