The right AI agent authentication tool depends on who the agent acts as: your own signed-in user, that user inside someone else's app, or itself. Identity platforms handle the first, token vaults and tool gateways handle the second, and per-action permission services sit across all three.
Picture the cheap version. Your agent needs to read a customer's calendar and post to their CRM, so somebody pastes an API key into an environment variable and moves on. Six months later that key still works, it can touch every account rather than one, nobody can say which agent run used it last, and the person who created it has left. One prompt injection, one leaked log line, and the agent's reach becomes an attacker's reach. You won't see that cost on the day you choose; it lands months later, usually at the worst moment.
So before any names, settle one question.
Who is the agent acting as?
There are three possible answers, and each one points to a different kind of tool: whose token the agent carries decides where it's stored, who can revoke it and what the logs can prove.
As your own user, inside your own product. Someone signs in to your app, then asks the assistant built into it to do things. The agent borrows that person's session and permissions. You need an identity platform: sign-in, sessions, roles, and now a way to issue the agent a narrower token than the human holds. If you're still choosing the base layer, the wider guide to the best authentication APIs covers sign-in itself; this page is about what happens once an agent joins in.
As your user, inside other people's products. The agent books meetings in Google Calendar, files tickets in Jira, reads a shared drive. Now you're running OAuth flows against dozens of third-party APIs, storing refresh tokens per user, and refreshing them before they expire. This is where token vaults and tool gateways earn their keep: they hold the per-user tokens so your agent code never sees a raw secret.
As itself. A background agent that reconciles invoices overnight has no human in the loop at run time. It needs a machine identity, short-lived credentials, and tight limits on what each run may do. Here the question shifts from login to authorisation: is this agent allowed to take this action on this record, right now?
A single product can easily need two of the three. A support assistant acts as the signed-in customer in your app and as a service account in your ticketing tool. That's fine, as long as you pick tools for each fork on purpose rather than stretching one tool across all of them.
Identity platforms that now cover agents
Auth0
A customer identity platform that now runs a separate product line for securing AI agents alongside ordinary app sign-in.
Descope
Teams that like building login journeys on a visual canvas use this identity platform, which treats customers and agents as two kinds of user.
WorkOS
Enterprise-flavoured auth and user management, with a single sign-in idea stretched to cover the agents working across your apps.
Clerk
For product teams wanting auth and user management in one place, this one adds agent authentication and help wiring up MCP servers.
Stytch
An authentication platform built with both humans and AI agents in mind, and now part of Twilio.
Token vaults, gateways and per-action control
Arcade
Sits between your agents and the systems they call, acting as a runtime that also takes care of the auth for each action.
Keycard
A small control plane for runtime agent auth, granting access one action at a time rather than handing out standing keys.
Scalekit
Gateway-style option: an agent kit and MCP gateway with ready connectors, keeping a separate token for every user.
Permit.io
Fine-grained authorisation as a service, for when the hard part is policy and audit rather than login.
Nango
Open-source and hosted integration platform that hands agents access to a very large catalogue of APIs with permissions managed for you.
What good agent auth looks like once it's running
Whichever fork you're on, a safe setup and a risky one differ in the same few ways, and they're worth checking before you look at any feature list.
Start with how the agent gets its access. It should hold its own token, derived from the user's consent and carrying fewer scopes than the user has, rather than a copy of the user's session. If a tool's quick-start has you pass the human's session token straight to the model's tool layer, treat that as a demo, not a design.
Then look at lifespan. A token that expires in minutes and is refreshed by the platform beats one that lives for months, but ask the awkward follow-up: how does refresh work when the user is offline? That's exactly when background agents run.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Revocation is where cheap setups fall apart. When a user disconnects an integration, or you pull the plug on an agent that's misbehaving, its access should end in one place without logging the user out of their own account.
And you'll want a record you can question later: which agent did what, for whom, under which grant. Plain sign-in logs can't answer that. Per-action records can.
The MCP servers most teams plug in today make this sharper, because a single agent may now reach ten tools in one conversation. Each of those tool calls is a moment where a scope can be too wide.
How the bill is worked out
Agent auth is priced in quite different units, and the one you pick changes which way your costs move as you grow.
Identity platforms price by the people using your product: monthly active users, with the agents those users run counted alongside them on some pages. That suits a product where each human brings one assistant along. Runtimes and gateways price by what the agent does instead: tool calls, transactions, or connections plus compute and data transfer. Activity-based billing is kinder when you have few users with busy agents, and harsher when a loop goes wrong and fires the same call thousands of times.
Free plans are easy to find in this category, so you can build and test the revoke rehearsal below without paying anything. Read the limit on each one, though: some free tiers stop at a user count, others at a monthly number of calls or transactions, and a hard cap stops your agent rather than sending a bigger invoice.
Once you outgrow the free tier, the first paid monthly plan on the identity and gateway pricing pages re-read this month runs from tens of dollars to several hundred dollars a month. That's the entry ticket, not the bill. On top of it come the add-ons that move the number: extra users beyond the plan's allowance, extra transactions or tool calls, and enterprise single sign-on, which some of these pages charge per customer connection. If you sell to larger companies, that last one can outweigh everything else.
Before you commit, model a bad day rather than an average one: a runaway agent, a retry storm, a customer who connects every integration you offer. Activity-metered tools can turn that day into a surprise; per-user tools mostly won't notice it.
Run a revoke rehearsal before you choose
Feature pages won't show you the difference that matters, so test the one thing that separates them: what happens when access has to stop. Give it an afternoon with your two strongest candidates.
Wire up a single real integration, such as a calendar or a ticketing tool, with a test user. Let your agent take three actions through the tool: read something, create something, delete something it created. Then, mid-session, revoke that user's consent from the tool's dashboard and ask the agent to try again. Note how long the old token keeps working, what error the agent sees, and whether your logs show the three earlier actions against the right user and the right agent.
Finally, swap the test user for a second one and confirm the agent can't see the first user's data. If any of those steps needs a support ticket or a code change in the vendor's library, you've learned more than a sales call would tell you.
Do you make authentication, token vault or authorisation tools for AI agents? Tool makers can ask to be considered for a featured place on this list.
Mistakes that turn into incidents
- Giving the agent a service account with admin rights because scoping was fiddly. It's the long-lived key from the opening, wearing a nicer name.
- Storing refresh tokens in your own database to save a step, then discovering you've built a token vault without the encryption, rotation or audit a dedicated one brings.
- Treating sign-in and authorisation as one problem. Knowing who the agent is says nothing about whether it may refund this order.
- Letting the model see raw credentials in its context. Anything in the prompt can end up in a reply, a log or a trace.
- Skipping human approval on irreversible actions. Payments, deletions and outbound emails deserve a confirm step, whatever the tool.
If you're building on an agent framework, check how it passes credentials to tools before adding a vendor. The guide to the Claude Agent SDK walks through its permission checks and the pattern of injecting credentials through a proxy so the agent never sees them.
Which type to pick
You're adding an assistant to your own SaaS product. Start with an identity platform that already issues agent tokens. You'll keep one user directory, and agents inherit your existing roles.
Your agent mostly works inside other companies' apps. Pick a token vault or tool gateway with ready connectors for the APIs you need. Count the connectors you'd otherwise build yourself; that's the real saving.
Your agents run unattended and touch money or customer records. Add a per-action authorisation layer on top of whatever handles identity. It's the piece that lets you say no to one action without switching the whole agent off.
You're early and unsure. Choose by the unit you'll be billed in, not the feature list. If your users will each run many agent actions, per-user pricing is the safer bet; if you'll have few users with occasional actions, activity-based pricing costs less.
Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.


