Arcade vs Composio vs Pipedream Connect vs Nango: Who Holds the User’s Token

One of these four bills you for OAuth challenges as its own line item, separately from tool executions. That is the most honest pricing page in the category, because the connector catalog all four advertise is the part you will outgrow and the token vault none of them market is the part you would hate to build. And the decision you genuinely cannot retrofit is not on any pricing page at all: whose name appears on the consent screen your users click.

At a glance

Four platforms that stand between your agent and a user's Gmail, Salesforce or Linear account. All four hold tokens. They disagree about what to charge for and whose product it looks like.

Platform	Catalog, as it counts it	What the meter runs on	Where the tokens can sit
Arcade	7,500+ prebuilt tools across 81 MCP servers, grown from around 60 services in a year.	User authorization challenges and tool executions, priced separately, with a split between standard and "pro" calls.	Hosted, or a self-hosted Engine and workers — including disconnected deployments.
Composio	Roughly 1,000 third-party applications behind a single hosted MCP URL.	Tool calls. Free tier around 20k calls a month, paid plans from about $29.	Hosted, with an MIT-licensed SDK and a self-hostable MCP server.
Pipedream Connect	3,000+ APIs and 10,000+ tools through one MCP endpoint — the widest published catalog here.	Credits, where a credit is 30 seconds of execution at 256 MB and scales with memory, plus a per-external-user meter.	Managed. Credentials are held server-side by Pipedream.
Nango	900+ APIs with 700+ pre-built connectors, plus MCP Auth since February 2026 for authorising third-party MCP servers.	Connections and requests — the free tier is 10 connections and 100k requests; Starter is around $50 for 20 and 200k.	Hosted, a free open-source self-hosted edition for auth and the proxy, or full enterprise self-hosting.

Published pricing as of August 2026; every number in this table is the kind that moves. The structure of each meter moves much more slowly, and the structure is what you should be choosing on.

The catalog numbers are not the same number

Four vendors, four denominators. The bars are not comparable, which is the finding rather than a caveat.

Pipedream counts tools and APIs. Arcade counts tools and MCP servers. Composio counts applications. Nango counts APIs and connectors. A "tool" is one callable operation; an "app" or "API" is a whole integration that may expose fifty of them or three. Ranking these platforms by their headline figure compares a count of verbs against a count of nouns.

The number you actually need is a boolean, and you can compute it in an afternoon: do the three or four integrations my customers keep asking for exist here, and are the specific operations I need among the ones implemented? Every catalog in this category is deep on the same dozen SaaS products and thin on the long tail, and the long tail is where your differentiation lives. Depth per integration is also the axis no vendor publishes, because it is the one that would be embarrassing.

There is a second reason not to weight the catalog heavily. Adding an integration is work you can do; every one of these platforms supports custom tools, and Arcade ships a framework specifically for building them. Reproducing a production token vault is work you cannot casually do. Buying on catalog size optimises the reversible half of the decision.

The vault is the product

The property worth paying for is the dotted line: the token is injected at execution and never enters the model's context.

Strip the marketing and all four sell the same core mechanic. Your agent asks for a tool by name and a user ID. The platform looks up an encrypted, per-user, per-provider credential; if there is not a valid one, it puts the user through an OAuth challenge and remembers the result; then it injects the token into the outbound call at execution time and hands your model only the response. The model never sees the secret, which means a prompt injection in a returned document cannot exfiltrate a credential that was never in the context window.

Written out, that mechanic is unglamorous. Building it is not. It is per-provider OAuth quirks, refresh races between concurrent tool calls, rotation, scope-mismatch recovery, revocation, and a storage layer that has to survive a security review. This is why scoped credentials is the topic teams reliably underestimate and then rebuild badly under deadline.

Arcade's pricing is the clearest statement of this in the market: it meters user authorization challenges as a line item, distinct from tool executions, on every tier down to the free one. Whatever you think of the rate, the packaging is honest — the vault is a product with its own unit, not a free accessory to a catalog. The other three bundle it, which makes it easier to sell and easier for a buyer to undervalue.

Whose name is on the consent screen

This is the section to read if you skip the rest. When a user of your product clicks "Connect Gmail", an OAuth consent screen appears, and it says a name. On the default path with a hosted platform, that name is the platform's, not yours: point an agent at Composio's hosted MCP URL and your end users see Composio's brand in the flow unless you supply your own client credentials. Nango's pitch runs the other way: a white-label Connect UI across its catalog, designed so your brand is the one on screen.

Three reasons this outranks catalog size:

It is the least reversible thing in the stack. Tokens are issued against a specific OAuth client. Changing which client requests them means every user re-authorises. A re-consent campaign across a live user base is not a migration task, it is a conversion funnel, and some fraction of users never come back.
It is a trust event, at the worst moment. A screen that asks for access to a customer's mailbox in the name of a vendor they have never heard of is the highest-friction step in your onboarding, and it fails silently — enterprise users just close the tab.
Some approvals are literally about the name. An IT admin allow-listing an OAuth app approves a specific client. If that client is your integration vendor rather than you, the review you passed is not the review you thought you passed, and it has to be redone if you ever change vendors.

Bringing your own OAuth client IDs is supported across these platforms and is the correct default for anything customer-facing. Do it at adoption. It costs a day of app registrations up front, and it is the one change in this comparison that gets more expensive every week you defer it.

The meter shapes the agent you build

Every one of these meters is defensible. Each one taxes a different kind of agent.

Pricing units are usually treated as a procurement detail. In this category they are an architectural constraint, because agents differ enormously in shape and each meter charges for a different one.

Per tool call — Composio's model — is simple and it taxes chattiness. A ReAct-style loop that lists, filters, then reads is three calls where a purpose-built tool is one; retries and self-corrections are billable; and the cheapest thing you can do under this meter is exactly what makes agents better anyway, which is to design coarser tools. Credits denominated in execution seconds and memory — Pipedream's — taxes duration instead, so a fast API call is nearly free and anything that waits, paginates or processes a payload costs real money; the per-external-user meter on top makes wide, shallow consumer bases the expensive case. Connections and requests — Nango's — makes the number of connected accounts the primary axis, which is generous to an internal agent hammering ten integrations and unforgiving to a consumer product with fifty thousand lightly-used connections. Challenges plus executions — Arcade's — separates the cost of onboarding a user from the cost of serving them, and it is the only one that makes a re-authorisation storm visible as a cost rather than as a mystery.

Take your own numbers — tool calls per task, tasks per user per month, connected accounts — and price the same workload under all four. The spread between the cheapest and dearest for a given shape is routinely larger than the spread between the platforms' list prices, and the cheapest one for a chatty internal agent is often the dearest one for a consumer app.

Where the tokens are allowed to sit

The fourth axis is the one that decides the shortlist in a regulated environment, and it inverts the catalog ranking. Pipedream, with the widest catalog, is managed-only and holds credentials server-side. Arcade's Engine and workers can be self-hosted, including in disconnected environments, with tenant isolation, audit logs, RBAC and SSO on the enterprise tier. Nango publishes a free open-source self-hosted edition covering auth and the API proxy, with full-platform self-hosting for enterprise. Composio's SDK is MIT-licensed with a self-hostable MCP server, though the managed catalog is the path everything is optimised for.

Notice the trade: the two platforms most willing to let the vault live in your infrastructure are the two with the smaller or more narrowly-counted catalogs. That is not a coincidence. A hosted catalog of thousands of integrations is exactly the asset that becomes harder to ship into someone else's VPC. If data residency is a hard requirement, decide that first — it removes options faster than any other criterion here.

When to pick which

Situation	Lean	Because
Customer-facing product, users connect their own accounts	Nango, or Arcade with your own client IDs	White-label consent is a first-class feature rather than a workaround, and the consent screen is the irreversible decision.
Internal agent, breadth matters more than branding	Pipedream Connect	The widest catalog by a wide margin, and nobody internal cares whose name is on the OAuth screen.
MCP-native stack, per-user authorisation is the whole problem	Arcade	Built as an MCP runtime around authorised tool calling, with auth priced as its own unit and a framework for custom tools.
You need to reshape tool schemas the model sees	Composio	Code-based modifiers rename fields, hide parameters and simplify responses — the practical fix for a tool surface designed for humans.
Regulated, air-gapped, or tokens may not leave your VPC	Arcade or Nango, self-hosted	Both ship real self-hosted credential layers; the fully managed options cannot meet the requirement at any price.
Three integrations, one of them exotic	None of them, yet	Two OAuth flows and a token table is a week. Adopt a platform when the count grows or when per-user auth becomes the hard part.

FAQ

Which has the biggest catalog?

Pipedream Connect, at 3,000+ APIs and 10,000+ tools, by any reading. But it counts tools while Composio counts applications and Nango counts APIs, so the headline figures are not directly comparable — and depth on the specific integrations you need matters more than the total.

Can I keep my own branding on the OAuth screen?

Yes, by registering your own OAuth client with each provider and configuring the platform to use it. Nango makes white-label the default path. Do this before you have users: changing the OAuth client later forces every existing user to re-authorise.

Does the model ever see the access token?

On all four, no — the credential is injected server-side at execution time and only the result returns to the model. That is the main security argument for a platform over hand-rolled tools, where the token often ends up in a header your agent code assembles inside the loop.

Can I self-host the credential store?

Arcade and Nango both offer real self-hosted deployments, Arcade including disconnected environments and Nango with a free open-source edition covering auth and the proxy. Composio publishes an MIT SDK and self-hostable MCP server. Pipedream Connect is managed.

Which is cheapest?

It depends on the shape of your agent, not the list price. Per-call pricing punishes chatty loops, execution-second credits punish long or heavy calls, and per-connection pricing punishes a wide consumer base. Price your own call-and-account volumes under all four before comparing headline tiers.

Do I need one of these at all?

Not for two or three integrations — that is a week of OAuth work. The tipping point is per-user authorisation across many providers, where refresh, rotation and revocation become a standing engineering cost rather than a one-off.