Composio vs Nango vs Arcade.dev vs StackOne vs Merge: Which Agent-to-App Connectivity Platform to Build On

Start by crossing two off the list

Of the seven platforms usually named in this category, two no longer belong in a buy-vs-build comparison.

Smithery is gone as an independent option. Arcade.dev acquired it on August 5, 2026 — Smithery's registry and one-click MCP hosting are now being folded into Arcade's runtime, with co-founder Anirudh Kamath joining Arcade's team (announcement). The product still runs at smithery.ai today, but you'd be building on Arcade's roadmap, not a separate company's. Evaluate Arcade.dev instead.

mcp.run has left the self-serve market. The domain now redirects straight to a rebranded product called Turbo MCP — a self-hosted, enterprise-only "MCP gateway and management platform" with no visible pricing, no free tier, and every call to action routed to "Book a Demo" (turbomcp.ai). If you were evaluating mcp.run as a lightweight way to try MCP hosting without a sales call, that door has closed; what's left is a governance product aimed at IT buyers doing procurement, not developers doing a Tuesday-afternoon integration.

That leaves five real contenders: Composio, Nango, Arcade.dev, StackOne, and Merge. They don't compete on one axis — they meter usage in three fundamentally different ways, and which one is cheap for you depends entirely on which shape your workload takes.

Three pricing models, not one

Composio and Arcade.dev price per tool call. Every time your agent actually executes an action — sends the email, updates the CRM record — that's a metered unit. Composio's Free plan includes 100,000 tool calls a month with unlimited connected accounts; past that, its Scale plan is $29/month plus $0.0003 per call (composio.dev/pricing). Arcade's Free plan includes 2,000 tool calls and 2,000 auth events a month; its Team plan is $25/month plus $0.01 per tool call and $0.10 per auth event (arcade.dev/pricing).

That's not a rounding difference. At 50,000 tool calls a month with 50 connected accounts — a plausible mid-size internal agent — Composio stays inside its free tier at $0/month. Arcade's Team plan, doing the same volume, comes to roughly $25 platform fee + $500 in tool-call usage + $5 in auth events ≈ $530/month, because Arcade's per-call rate is over 30x Composio's. If your agent is going to make a lot of tool calls, that rate difference compounds fast; if you're only occasionally invoking a handful of high-value actions, it barely matters. Check your own expected call volume against both cards before assuming either is "the cheap one" — Arcade's free tier and lower connector count can still make sense for a low-volume prototype.

Nango prices by connection plus compute time and data transferred, not by call. Its Free plan is capped at 10 connections, 10 compute-hours, and 10GB of data transfer a month; past that, its Pay-as-you-go plan runs $0.29 per connection, $0.72 per compute-hour, and $0.50 per GB, with a $50/month floor (nango.dev/pricing). This rewards workloads with many connected accounts that don't hammer the API constantly — 50 connections costs about $11.60/month in connection fees alone, before compute — and penalizes agents that run long or frequent jobs against each connection, since compute-hours accrue regardless of how many discrete "calls" happened inside them.

StackOne and Merge price by seat or linked account, closer to traditional SaaS. StackOne's free Starter tier includes 1,000 credits per seat per month (one tool call or API call = 1 credit); its Team tier starts at $600/month for 5,000 credits per seat (stackone.com/pricing). Merge's Launch plan gives you 3 free production Linked Accounts, then $650/month for up to 10, with Professional and Enterprise both contract-priced with no published rate card (merge.dev/pricing/unified). Neither meters raw agent activity the way Composio or Arcade do — they meter how many end users or accounts you're serving, which fits a product you're embedding integrations into far better than it fits a single internal agent making a lot of calls.

The practical takeaway: don't compare sticker prices. Compare the metering unit against your actual architecture. An internal agent hitting a handful of tools a lot wants per-call pricing with a cheap rate (Composio). A platform onboarding many customers, each with a few connected accounts, wants per-seat or per-connection pricing (StackOne, Merge, Nango).

Licensing: only one of these is actually open source

Composio ships its core platform under MIT and has the community to show for it — 30,193 GitHub stars and a commit pushed today (github.com/ComposioHQ/composio). Arcade's MCP framework repo is also MIT, though smaller — 1,025 stars, pushed within the last two days (github.com/ArcadeAI/arcade-mcp).

Nango publishes its full source at 12,159 stars, but under the Elastic License 2.0, not an OSI-approved open-source license — it explicitly bars you from offering the software "to third parties as a hosted or managed service" that reproduces its functionality (Nango LICENSE). You can read it, self-host it, and modify it; you can't build a competing hosted Nango on top of it. That's a meaningfully different promise than MIT, and worth knowing before you architect around "we can always self-host if pricing changes."

StackOne and Merge don't compete on this axis at all: StackOne's only public repos are small SDKs, and Merge has no public repository. Both are closed-source SaaS: what you're buying is the hosted service, full stop, with no self-hosting escape hatch if terms change later.

Who should build on what

Default to Composio if you're building an agent that needs a large catalog of pre-built tool integrations, a generous free tier to prototype against, and the biggest, most actively developed open-source base in the category. At 1,500+ toolkits, an MIT license, and a free tier that covers real production volume (100K tool calls/month, unlimited connections), it's the hardest of the five to argue against for a general-purpose agent.

Choose Nango if your workload is connection-heavy rather than call-heavy — you're managing hundreds of customer integrations that don't get hit constantly — or if you specifically want to read and audit the source even though you can't legally re-host it as a competing service.

Choose Arcade.dev if you're already committed to MCP as your protocol and want a runtime built around governance and audit from day one, and if your expected tool-call volume is low enough that its steeper per-call rate doesn't matter. The Smithery acquisition also makes it the platform to watch for registry/discovery features over the next few months.

Choose StackOne if you're an enterprise deploying an internal agent gateway across many employees and need SOC 2, HIPAA, and GDPR compliance included on every tier, including the free one — that's not something the other four commit to at the entry tier.

Choose Merge only if what you actually need is a normalized data model across HR, ATS, CRM, and accounting systems — Merge's Common Model schema, not raw tool-call execution — and you're comfortable negotiating a contract past the first 3 linked accounts.

Skip mcp.run/Turbo MCP unless you're already an enterprise buyer prepared to go through a sales process for a self-hosted gateway; there's no self-serve tier left to evaluate. Skip Smithery as a standalone bet — it's Arcade's product now.