Agent auth platforms compared, 2026: Arcade, Composio, Nango, Scalekit, Pipedream, Auth0, Paragon, Merge and Anlyon

Every platform below connects your AI agent to your users' own accounts and keeps the tokens out of your app. They differ on what happens after that: whether a write is limited per user, approved by that user on the exact request, and recorded with a receipt you can check.

If you need hundreds of integrations on day one, pick one of the large catalogues. If your agent writes in a handful of your users' tools, and a wrong email, a closed deal or a duplicate record there is your problem, that second half is what Anlyon is for. Many teams will want both.

Every fact about another product on this page comes from that vendor's own docs or site, with the date we read it. Where we looked for something and did not find it, the cell says "not found" and nothing stronger. Each vendor has its own page with sources under every cell, linked below.

The short answer

What does an agent auth platform have to do?

Connecting the account is step one. When your agent goes from reading a user's account to writing in it, three questions arrive in the same week:

1. Whose account is it? Each user connects their own HubSpot, Slack or Linear. Someone has to run consent, store tokens, refresh them, revoke them and delete them when the user leaves. Every platform here does most of this. See what per-user OAuth takes to build.
2. How much may the agent do there? "Each user's assistant may make 20 CRM changes a day" is a per-user limit on writes, not a rate limit on requests and not a token budget. See per-user limits.
3. Did the user say yes, to this exact request? A user approving "send an email" is not the same as approving this email, to these people. See approvals your users decide in your app.

Then a fourth, after the fact: what did the agent do for this user, and did it land? See how to query a user's actions and receipts.

The full comparison

Checked 2026-11-02, except Arcade and Composio (2026-10-24 for per-user access, 2026-10-03 for catalogue counts). Read each vendor page below for the source of every cell.

"Not found" means we looked in that vendor's docs on the date above and did not find it. It does not mean the product cannot do it. If one of these cells has gone stale, tell us at support@anlyon.com and we will correct it.

What Anlyon does differently

Anlyon makes the call itself. Your agent names an action, your server names the user, and Anlyon holds that user's token, applies the user's limit and approval, calls the provider and returns a receipt.

Without Anlyon   agent -> the user's token in your process -> provider API
With Anlyon      agent -> named action -> Anlyon -> provider API

Because Anlyon owns the call, it can enforce things a token broker cannot:

• A limit per user, counted before the call. One rule, such as 20 HubSpot changes a day, is counted for each user separately. The 21st change is refused before it reaches HubSpot, and no other user is affected.
• The user's yes, on the exact request. A policy with decidedBy: "end_user" sends the approval to that user, in your app. The approval is bound to a SHA-256 digest of the request, and Anlyon recomputes it before dispatch. What the user approved is what runs.
• A receipt with a grade. confirmed means a read-back of the provider's own record matched. acknowledged means the provider accepted it and nothing read it back. unknown means no usable response, and Anlyon does not send it again.
• The server names the user. anlyon.forUser(session.user.id) binds the client. The tools the model sees have no user field, and the client cannot be switched. See can a prompt make your agent act as another user?
• Connect, then confirm. A connect link works once, for ten minutes, and your server confirms the connection for the user in its own session, so a forwarded link cannot attach someone else's account.
• Erase in one call. endUsers.erase(userId) revokes and deletes connections, cancels pending approvals and clears limit counters. See deleting a user from your AI agent.

import { Client } from '@anlyonhq/sdk';

const anlyon = new Client({ apiKey: process.env.ANLYON_API_KEY! });

// Your session names the user. The model never does.
const user = anlyon.forUser(session.user.id);
const result = await user.actions.invoke('hubspot-update-deal-stage', {
  dealId: '1234567',
  dealstage: 'closedwon', // a policy can send this to the rep to approve first
});

Where Anlyon is not the right fit today: the list is short (sign-in for HubSpot, Slack, Linear, GitHub and Notion, plus about 27 integrations by the user's own key such as Attio, Asana, Jira, Zendesk and Pipedrive; Gmail, Google Calendar and Salesforce are on the way), end users are supported over the HTTP API and both SDKs but not yet over MCP or the CLI.

Each comparison in detail

• Arcade and Composio
• Nango
• Scalekit AgentKit
• Pipedream Connect
• Auth0 for AI Agents and Token Vault
• Paragon ActionKit
• Merge Agent Handler
• Anlyon vs building it yourself

Using more than one

The common setup is a large catalogue for reads and the long tail, and Anlyon for the writes that need a limit, an approval and a receipt. A user connects each product separately; connections and tokens are not shared between them. Start with how to let your AI agent act in your users' accounts, or open the console and send your first user a connect link.