Arcade

What it is

Arcade is the authorization and execution layer between an agent and the systems it writes to. Where Composio competes on catalogue breadth at the lowest price per call, Arcade competes on a different question: which human authorized this write, and can you prove it six months later. When a tool call needs an OAuth scope the calling user has not granted, the Arcade Engine interrupts the run, walks that user through the provider’s consent flow, and resumes — so the token in play belongs to the person making the request rather than to one shared platform key. It reads your existing identity provider (Okta, Entra) instead of replacing it, and evaluates per-action authorization at call time.

The company raised a $60M Series A in June 2026 led by SYN Ventures, with strategic investment from Morgan Stanley and Wipro, bringing total funding to $72M including a $12M seed. It authored the specification enhancement proposal behind MCP’s URL elicitation flow, developed with Anthropic and accepted into the MCP spec in November 2025. On 5 August 2026 it acquired Smithery, the public MCP registry and hosting platform, on undisclosed terms; Smithery co-founder Anirudh Kamath joined the team.

Why it shows up in agent stacks

One platform OAuth client is one blast radius. Composio disclosed on 21 May 2026 that attackers exfiltrated roughly 5,241 API keys and 5,001 GitHub OAuth tokens from exactly that kind of central store. Arcade’s per-user model shrinks the same incident to the grants of whichever individual accounts were reached, and leaves an action log that names the authorizing human — which is the artifact a compliance reviewer asks for and a shared service account cannot produce.

Read the catalogue numbers carefully. Arcade claimed more than 8,000 agent-optimized tools in June 2026, but the docs list roughly 60 named MCP servers across eight categories, and splits them into Optimized (Arcade-built) and Starter (generated API wrappers). Tools are individual functions; servers are what you actually integrate against.

Pricing reality

• Free — $0, with 2,000 tool calls and 2,000 auth events per month, managed cloud hosting, and support through Discord and GitHub.
• Team — $25/mo platform fee, then $0.01 per tool call and $0.10 per auth event, with next-business-day email support.
• Enterprise — custom, sold as annual bundles of auth events and tool calls at discounted rates. This tier carries SSO, RBAC, audit logs, private registry access, self-managed deployment on Helm/Kubernetes or the Azure Marketplace managed app, and 24/7 SLA-backed support.

There is no per-seat licensing. The real band: an internal agent serving 300 users at 25K tool calls a month lands near $365 in its first month ($25 platform, $250 in calls, about $90 to connect 900 user-provider pairs), then near $275 steady-state. At 500K tool calls a month the Team list price is $5,025 — the volume where you stop paying list and negotiate a bundle.

Per call, $0.01 against Composio’s $0.0003 is a 33x gap. That is the price of per-user tokens and per-action audit. If nobody will ever ask which human authorized a write, you are paying two orders of magnitude for a property you do not use.

Auth events scale with your user count, not your usage. Every new user-provider connection bills $0.10, and re-authorization after a token expiry, a scope change, or a revoked grant bills again. A 5,000-seat rollout across four providers is 20,000 auth events before the agent makes a single tool call.

Best for

The platform or security engineer shipping an agent that writes to a system of record on behalf of many named end users, where the audit requirement is “prove which human authorized this action.” The scoped case where it wins outright: a RevOps or Customer Success team putting an agent in front of several hundred reps that updates Salesforce or HubSpot and posts to Slack as each rep, with a per-user trail that survives a compliance review.

Not for

A single-tenant internal agent running as one service account at high call volume. If the same three tokens serve every request and no reviewer will ever ask whose authority a write carried, the 33x per-call premium buys nothing. It is also the wrong layer for a non-engineer — this is an SDK, a CLI, and an API, not a canvas. And for a legal-ops or recruiting team whose data-residency rules forbid a third party holding tokens at all, the self-managed deployments answer the objection but move you to Enterprise pricing to get there.

Versus the alternatives

• Composio — roughly 1,000 integrations at $0.0003 per call, the cheapest authenticated tool call in the segment. Pick it when call volume dominates the bill and a single platform key is custody your security review will accept. Its May 2026 breach is the argument against that acceptance.
• Zapier MCP — the largest install base in the category at 9,000+ apps, included on every Zapier plan. Pick it when an ops generalist is doing the wiring, monthly volume stays under about 1,000 tool calls, and Zapier is already on the invoice. Task-based billing works out near $0.069 per call, so it inverts the moment volume grows.
• Runlayer — the fastest-growing entrant in agent governance: an $11M seed in November 2025 followed by a $30M Series A in June 2026 led by Felicis, with 18,000 pre-built MCP connectors, approval workflows, and scanning for prompt injection and data exfiltration; its Pro plan is $499/mo for 25 registered MCP servers. Pick it when the problem is shadow-MCP discovery and inspecting traffic from servers you did not build. Arcade wins on identity, Runlayer on inbound threat inspection; teams with both problems buy both.

If none of these fit, the honest fallback below about five integrations and one service account is writing direct API clients and skipping the layer.

Watch-outs

• The homepage says “SOC 2 compliant” without naming the type, and lists no ISO 27001 or HIPAA position. Guard: ask for the report, its type, the audit period, and the most recent pen-test summary in writing before the security review starts, and treat the marketing line as unverified until the document is in hand — a legal-ops or recruiting deployment touching contracts or candidate PII will not clear review on a homepage claim.
• Auth-event billing is driven by headcount, so it spikes at rollout and again at every forced reconnection. Guard: model it as users × providers × (1 + your annual re-auth rate), set a spend alarm on that line specifically, and negotiate auth events into the Enterprise bundle before the rollout rather than after the first invoice.
• The Free tier is a demo, not a pilot. 2,000 tool calls is one afternoon of a chatty agent, and the ceiling arrives without warning mid-evaluation. Guard: run any pilot that carries a go/no-go decision on Team with an explicit budget cap, so the evaluation ends on evidence rather than on a quota.
• Starter servers are generated API wrappers, not hand-built integrations, and the quality gap is real. Arcade’s own index of 43,400+ MCP servers and 219,069 tools found that 0.5% earned an A grade. Guard: confirm the specific server you depend on is Optimized before you commit, and test the three tool calls your workflow relies on against live data during the trial rather than reading the catalogue entry.