IANA-AP

Intent Agent Native Authorization
for Agentic Profiles

Traditional authorization asks if a request is allowed. Agentic systems need to ask something more: does this request still belong to the plan you approved? IANA-AP is a runtime authorization framework for AI agents. Before any tool runs, the agent discovers what authorization is required, computes its intent, and surfaces the full plan for review. You approve it inline with a Passkey, one cryptographic gesture that binds every action to the declared scope. If a call falls outside the approved intent, it does not execute.

The Framework

IANA-AP

Intent Agent Native Authorization for Agentic Profiles is a runtime authorization framework that composes OAuth 2.0, OpenID AuthZEN, WebAuthn/FIDO2 Passkeys, and MCP, plus TwoGenIdentity agentic frameworks, rather than replacing them. Unlike deterministic software, an agent only discovers what authorization it needs once it understands the request. IANA-AP closes that gap: the agent declares its full plan, the user reviews and approves it with a single Passkey gesture, and a cryptographically bound token is issued. Every subsequent tool call is verified against that signed intent. Agents earn the right to act. They do not assume it.

How It Works

The Four Phases of IANA-AP

IANA-AP replaces implicit trust with a strict authorization lifecycle. Every agent action passes through four phases, from discovering what's required to enforcing it at runtime, each one a verifiable cryptographic boundary.

Discover Required Authorization

The agent reads the service's Agentic Profile, a machine-readable definition that declares what authorization each tool requires.

Phase 1

Compute Authorization Intent

Using the discovered requirements and the user's prompt, the agent works out its planned operations and assembles them into a proposed intent.

Phase 2

Review and Approve

The full plan is surfaced natively inside the agent interface, no browser redirect. The user approves with a single phishing-resistant Passkey gesture.

Phase 3

Runtime Enforcement

Every tool call is intercepted and checked against the approved intent before it runs. Anything outside the signed scope is blocked.

Phase 4

See It In Practice

Inside an Intent Agent Native Authorization Review

This is what the agent surfaces before it acts: the full plan, waiting on your Passkey. Nothing executes until you approve it. Here's what happens in each phase along the way.

Phase 1

Discover Required Authorization

  • Reads the Agentic Profile

    Before doing anything, the agent reads the service's Agentic Profile via the MCP Server Card or tools/list response.

  • Only tools that opt in are pre-authorized

    Each tool declares participation with an authorization mapping. Tools without one fall back to just-in-time approval instead.

Supported Agentic Profile Targets

MCP A2A OpenAPI

Authorization Mapping in Agentic MCP Profile

Each tool opts in to IANA-AP with an x-authz-mapping that declares the authorization required to invoke it.

Phase 2

Compute Authorization Intent

  • Matches your prompt to what's needed

    The agent maps your prompt against the mappings discovered in Phase 1 to work out exactly which actions, on which resources, are required.

  • Can plan, but can't act

    This phase can run as a dedicated Intent Agent: a constrained component that computes the plan but has no ability to invoke tools.

  • Produces the agent_intent claim

    Each planned tool call becomes one agent_intent entry, a structured claim the user will review and approve as a set in the next phase.

The SARC Model

S

Subject

A

Action

R

Resource

C

Context

Worked example, same plan shown in Phase 3:

Subject

alice, from the token

Action

tools/call

Resource

mcp-tool: disable_user

Context

the agent's identity

Phase 3

Review and Approve

  • Discovered and computed before you see it

    By the time this card appears, the agent has already read what's required and worked out exactly what it needs to do.

  • Nothing runs until you approve

    The full plan sits here, visible end to end. You can inspect every action before deciding.

  • One Passkey tap, every action covered

    A single gesture approves the whole plan, not one prompt per action.

  • You can approve some, deny others

    Approval isn't all-or-nothing. The Authorization Server can approve individual actions and deny the rest; the agent may only invoke tools whose entry was actually approved.

agent · intent review
Discovered
Computed
Review & Approve
Enforce

Execution Plan

3 actions awaiting your approval:

disable_user user: alice urn:twogenidentity:mcp
revoke_sessions user: alice urn:twogenidentity:mcp
audit_log_export resource: logs/2025-05 urn:twogenidentity:mcp

Awaiting Passkey approval to bind this plan

One gesture approves all 3 actions · nothing executes until then

Phase 4

Runtime Enforcement

  • Enforced exactly as approved

    Once you tap, only these actions can run. Anything outside the plan is blocked at runtime.

  • No silent scope creep

    Every call is checked against the signed plan again, every time. Approval never quietly expands.

  • The same SARC fields work with any PDP

    Whether the decision is made remotely via AuthZEN or locally with Cedar or OPA, the enforcement point extracts identical Subject/Action/Resource/Context fields. Only the evaluation mechanism changes.

Policy Decision Point: Remote or Local

Remote PDP

OpenID AuthZEN

Local PDP

Cedar or OPA

Enforced at Every Stage of the Journey

MCP / AI Gateway PEP API Gateway PEP

Put It All Together

The Full Communication Sequence

The four phases above happen as one round trip. Here's how a single request travels from prompt to result. Tap a step to see the detail.

1
User Agent

"Disable this user and export the audit log."

Live Demo

See IANA-AP in Action

Full end-to-end demo of an agent orchestrator with A2A, MCP, and MCP Apps. The agent discovers what's required, computes and surfaces its plan, the user approves it with a Passkey, and every call is cryptographically verified against the signed intent.

IANA-AP: Full Demo

Discover Agent reads the Agentic Profile to learn what each tool requires
Compute Agent computes its intent from the plan and surfaces it inline, no browser redirect
Review & Approve User approves with a device-bound Passkey, one gesture
Enforce Scope-bound token issued · every call verified against the signed intent

Composable Authorization

Combine Authorization Layers for Your Security Posture

Each authorization layer is independent and composable, like Lego blocks. Activate the layers that match the risk level of the operation. You are not locked into a single model. And within IANA-AP itself, no redesign is required: it works with the Authorization Server and policy engine you already run.

1

OAuth Security

Standard access control via OAuth and OpenID Connect. The baseline for every agent operation.

2

Intent Agent Native Authorization for Agentic Profiles

Pre-approve the full execution plan. Every action is cryptographically bound to the declared scope before the agent runs.

3

Just-In-Time Agent Native Authorization

Runtime step-up authorization for sensitive actions, backed by Passkeys and Human-in-the-Loop.

Why IANA-AP

A security-first approach to agentic execution. The question isn't just whether a request is allowed, it's whether it still belongs to the plan you approved. Agents earn the right to act. They do not assume it.

Plan Before Execute

The agent cannot act until the full execution plan has been reviewed and explicitly approved. No silent or implicit actions.

Scope-Bound Cryptographic Proof

Every action is verified against a signed intent token. Anything outside the approved scope is blocked before execution.

Checked Again at Runtime

If the agent's call doesn't match what was approved down to the resource and its arguments, it's rejected, even mid-session.

Passkey-Backed Approval

A Passkey gesture approves all planned actions, phishing-resistant, non-replayable, and tied to the user's registered identity.

No Browser Redirect

Intent review and Passkey approval happen inline inside any AI assistant or CLI agent. No context switch, no friction.

Built on Open Standards

Built on open standards and TwoGenIdentity agentic frameworks. Interoperable, portable, and free from vendor lock-in.

Composable by Design

Layer Intent Authorization on top of OAuth Security or combine it with Just-In-Time Authorization. Each layer is independent. Activate what your security posture requires.

Explore the Full Platform

IANA-AP is part of the TwoGenIdentity composable authorization stack. Explore the ANA framework for Just-In-Time runtime authorization, the IA+ IAM Platform that powers both, and the AuthZEN MCP Gateway that enforces policy at every agent request.