You’ve deployed AI agents into production. They’re making decisions, calling APIs, and moving money. But can you prove, right now, that every one of those actions happened under the right identity with the right permissions?
Probably not. And you’re not alone.
A 2025 survey by Enterprise Management Associates (EMA), conducted in conjunction with Ory, found that 79% of organizations lack written policies for governing AI agents despite having already pushed them into production. That’s not a gap. That’s a governance void.
Meanwhile, 72.1% of organizations prefer self-managed LLM deployment on premises or in private clouds, and that preference correlates with wanting self-managed identity infrastructure too (EMA, 2025).
By 2026, the momentum hit a new threshold: According to the 2026 EMA survey, geared toward the financial sector, 83% of large and 70% of medium-sized organizations had deployed AI agents in production, and 98% plan to use them.
The identity question isn’t theoretical anymore. This list evaluates platforms that give developers deployment flexibility for non-human agent identities across self-hosted, on-premises or in the private cloud, and fully managed SaaS. We focused on authentication, fine-grained authorization, and policy enforcement at the point where agents actually execute actions.
Methodology: How We Evaluated
We assessed these platforms against five criteria, with a use case focus on coding agents, machine-to-machine APIs, agentic payments and e-commerce, and runtime tool governance.
- Deployment flexibility: On-premises, private cloud, and fully managed SaaS options, plus API parity across deployment models.
- Agent-native authentication standards: Support for OAuth 2.0, OpenID Connect, and the Model Context Protocol.
- Fine-grained authorization: RBAC, ReBAC, ABAC, Zanzibar-style permissions, OPA, or Cedar support.
- Point-of-execution policy enforcement and observability: Policy checks where agents actually act, not just where they authenticate.
- Scale evidence and operational resilience: Production identity volumes, horizontal scaling, and global edge performance.
The 2025 EMA survey behind several of our data points captured 271 qualified responses from 1,256 total participants, drawing on IT, security, and IAM professionals at organizations with 500 or more employees across the US, UK, Canada, France, and Germany.
A quick note on our ranking approach: This is grounded in vendor documentation, industry surveys, and public case material. It’s not hands-on product testing. We’ve weighed the evidence that’s publicly verifiable, and we’ll flag where a claim comes from a vendor’s own materials rather than an independent audit.
Why Deployment Flexibility Is the Buying Criterion in 2027
Here’s the messy reality most engineering teams are staring at. Organizations report using an average of three IAM platforms, and 34% use four or more, fragmenting authentication and authorization across their AI agent fleet (EMA, 2025).
That sprawl isn’t accidental. It’s the residue of teams bolting identity onto agent projects one at a time, with whatever tool was closest at hand.
The identity model chaos runs deeper. 60.5% of organizations manage AI agents with a hybrid human/service account model, 23.6% treat them as pure service accounts, and 10% provision them as human identities (EMA, 2025). Three incompatible approaches inside the same enterprise.
What happens when a service-account agent needs to act on behalf of a human? Or when a human-provisioned agent exceeds what a human role should be allowed to do? Nobody has a consistent answer.
The tooling gap is equally stark. 76% of organizations have deployed tools to identify agents internally, but only 55% have documented policies to govern their behavior (EMA, 2025). Detection without governance means you can see the agents moving through your systems but can’t enforce what they’re allowed to do once they’re inside.
And here’s the kicker: most teams already know their current IAM stack isn’t ready. The same EMA survey found that 59.4% of organizations say no on security readiness, 62.0% on scale, 48.7% on compliance, and 61.6% on resiliency when asked whether their existing IAM can handle agents (EMA, 2025).
That’s a majority rejecting their own infrastructure on every dimension that matters.
Some of this frustration traces back to the all-in-one IAM dream. Only 24% of organizations have achieved a unified IAM platform, while 52.8% would prefer one (EMA, 2025). But all-in-one platforms tend to be cloud-based SaaS products. They don’t fit organizations that need on-premises or private cloud deployments. The preference is there. The product fit often isn’t.
The 2026 financial-services data sharpens the picture. Approximately 68% of organizations are looking to bring IAM deployment in-house, and 51% cite rising costs as their biggest IAM challenge (EMA, 2026). SaaS lock-in is starting to hurt.
Teams want control over their identity infrastructure, but they also want the operational simplicity they’ve gotten used to from managed services.
That tension is what makes composable IAM an increasingly attractive model. Instead of one monolithic platform, organizations can select and integrate separate components: authentication, fine-grained permissions, consent management, and auditability. The pieces fit together via APIs, giving SaaS-like operational simplicity without surrendering deployment control (EMA, 2026).
For engineering teams building agent-heavy applications in Ruby or other languages, this modular approach means identity logic can live inside your own services. You call headless identity APIs, handle auth decisions in your code, and keep the infrastructure wherever your compliance requirements demand it.
Developers who need versioned, well-documented standard libraries for building out those integrations often start with reference documentation like Ruby API documentation by version when mapping how identity SDK calls will interact with core language features. The composable model turns identity from a platform you rent into a capability you own.
That shift, from renting identity to owning it, is exactly what the next five platforms address in different ways.
1. Ory
Ory is an API-first, composable identity platform built for machine-scale agent workloads. At its core, Ory Hydra handles OAuth 2.0 and OpenID Connect at web scale, while Ory Keto delivers Zanzibar-style fine-grained permissions.
The broader suite includes Ory Polis, Ory OathKeeper, and Ory Talos, giving teams a modular stack they can assemble to their exact requirements. Critically, the same APIs are available across self-managed Ory Enterprise License deployments, on-premises or in cloud, and as fully managed SaaS platform via the Ory Network.
- Deployment parity across the board: Every Ory product and capability works in self-managed environments on premises or in the cloud through the Enterprise License, and as a fully managed SaaS platform through the Ory Network. The managed offering includes compliance support, multi-region deployment, zero-downtime upgrades, and 24/7 incident response. Self-managed teams operate Ory Hydra and Ory Keto on their own infrastructure, with full API parity. For teams evaluating Ory AI agent identity management, that parity means the authentication, authorization, and policy controls you prototype on the managed network carry over unchanged to a self-hosted deployment
- Ory Agent Security at the point of execution: This is Ory’s key differentiator, as it enforces policy where agents actually execute actions, not just at the authentication layer.
- Agent-native standards and fine-grained authorization: Ory implements OAuth 2.0, OpenID Connect, and the Model Context Protocol. It supports RBAC, ReBAC, and Zanzibar-style permissions through a headless, API-first design. The use case coverage spans coding agents, M2M APIs, agentic payments and e-commerce, runtime policy enforcement, and parameter-level authorization for live MCP tools. Integrations include the Anthropic Claude Agent SDK, Gemini CLI, AWS API Gateway, OpenAI Codex, and many others.
- Scale and production proof: Ory infrastructure manages almost 3.3 billion identities. It powers identity and authentication for OpenAI, handling 1.2 billion weekly users as of September 2026, including agentic flows. OpenAI adopted an Ory Enterprise License for Ory Hydra authorization. OpenAI engineering manager Benjamin Billings described the relationship directly: “OpenAI wanted a partner that could help enable our vision for owning our identity processes, data, and success. We have a lot of partners, and Ory is one of our best.” Gartner also lists Ory with a product presence in Access Management, transitioning toward Customer Identity and Access Management.
Best for: Teams that need one platform across self-hosted, on-premises or in the private cloud, or as fully managed SaaS, with point-of-execution agent policy enforcement at machine scale.
Less ideal if: You want a fully turnkey SaaS with a low-code identity UI, or you lack the engineering resources for a headless, API-first platform.
Ory’s strength is in being simultaneously deployable, scalable, transparent, and composable. It achieves trillion-scale performance through horizontal scaling and global edge performance. Ground-up observability, trace-level log correlation, and performance intelligence come built in rather than bolted on after deployment.
2. Keycloak
Keycloak is a self-hosted identity and access management server that’s earned its place in teams prioritizing standards-based identity federation. It’s battle-tested, widely deployed, and entirely self-managed.
- Supports OpenID Connect, OAuth 2.0, and SAML for broad federation coverage.
- Provides fine-grained authorization services managed directly from the admin console.
- Includes LDAP and Active Directory integration, identity brokering, and social login capabilities.
Best for: Teams that need a self-managed identity platform with broad federation support and direct control over their infrastructure.
Less ideal if: You need SaaS deployment parity, an MCP-aware agent policy layer, or point-of-execution enforcement for non-human identities out of the box.
Keycloak covers the self-managed identity fundamentals well. Keycloak provides robust support for OpenID Connect, OAuth 2.0, and SAML, as well as fine-grained authorization services managed from the admin console.
However, for agent-specific security and deployment flexibility across SaaS environments, additional tooling may be beneficial.
3. Permit.io
Permit.io positions itself as a fine-grained authorization layer rather than a complete identity provider. If you already have authentication sorted and need to lock down what agents can actually do, this is where Permit.io fits.
- Supports RBAC, ABAC, and ReBAC for humans, services, and agents.
- Offers a local policy decision point using OPA or Cedar that runs inside your VPC.
- Includes an MCP Gateway for agent tool calls and supports managed or self-hosted deployment.
Best for: Teams that already have authentication in place and need a dedicated authorization layer with local policy enforcement and MCP-aware agent tool controls.
Less ideal if: You need a full identity platform that covers authentication, lifecycle management, and point-of-execution security in one product.
Permit.io is a strong fit for authorization-first teams. But it works best when paired with an existing identity stack. It fills the authorization gap; it doesn’t replace the broader identity infrastructure.
4. Auth0 (Okta)
Auth0 brings a managed SaaS approach to AI identity, with capabilities aimed at B2B and commerce use cases. It’s the path of least resistance if you’re already in the Okta ecosystem.
- Centralized fine-grained authorization, audit trails, and agent-on-behalf-of-user access patterns.
- Fully managed SaaS delivery, reducing operational overhead for identity infrastructure.
Best for: Teams already standardized on Auth0 or Okta that want a managed SaaS path for agent-on-behalf-of-user flows.
Less ideal if: You need self-hosted or private cloud deployment flexibility, or enforcement at the point where agents execute actions outside the SaaS boundary.
Auth0 simplifies managed adoption and integrates cleanly into an existing Okta environment. But it’s less suited to organizations that are actively pulling IAM deployment in-house or that need direct agent execution control beyond what a SaaS perimeter can provide.
5. CyberArk (Idira by Palo Alto Networks)
CyberArk, now branded as Idira by Palo Alto Networks, approaches identity security from the privileged access and machine identity angle. It’s an enterprise-grade platform with a heavy focus on securing the identities that have the most destructive potential.
- Offered in both self-hosted and SaaS-delivered forms, so privileged access controls can stay on-premises, run in a vendor-managed cloud, or be split across the two.
- Secrets management for machine identities, covering the credentials, keys, and secrets that applications and automated workloads use.
- Integrates with the wider Palo Alto Networks security portfolio following the Idira rebrand.
Best for: Enterprises where privileged access management and machine identity security are the primary risk.
Less ideal if: You need a developer-friendly, API-first identity platform with full authentication, authorization, and MCP support across self-hosted and SaaS deployments.
CyberArk/Idira addresses enterprise machine identity risk effectively. But its focus is on the privileged access and machine identity security domain rather than on developer-centric agent identity workflows.
Final Verdict
For teams balancing self-hosted control with SaaS convenience, Ory stands out because it hits all four pressure points at once: deployment parity across models, Ory Agent Security at the point of execution, composable IAM architecture, and proven scale with almost 3.3 billion identities and OpenAI identity traffic at 1.2 billion weekly users as of September 2026. That combination is rare.
Keycloak fits well if your world is strictly self-managed federation and you’re willing to layer agent-specific tooling on top. Permit.io makes sense when authorization is your missing piece and you already have authentication in place.
Auth0 is a natural step for teams already committed to the Okta ecosystem and comfortable with a managed SaaS boundary. CyberArk/Idira addresses the privileged and machine identity security domain for enterprises where that risk profile dominates.
Your decision should turn on whether you need point-of-execution enforcement or are comfortable with perimeter-level controls, and whether deployment flexibility matters enough to influence your identity architecture.
The agent governance gap isn’t closing on its own. The platform you pick determines whether you close it now or patch it later.