Join our Newsletter — 33% off our NHI Course
Home› Guides› Enterprise-Managed Authorisation for AI Agents (ID-JAG and Cross…
Guide Agent Identity & Access

Enterprise-Managed Authorisation for AI Agents (ID-JAG and Cross App Access): Guide for Identity Teams

PDF for newsletter subscribers · free to read online

← All guides
By NHI Mgmt Group, edited by Lalit Choda Updated 27 October 2026 13 min read
On this page

Every AI assistant that connects to an MCP server or SaaS API through a per-user OAuth consent screen creates another long-lived grant that the identity team did not approve and cannot revoke centrally. Multiply that by dozens of MCP servers and thousands of employees and the result is consent sprawl: tokens held inside agent platforms, no joiner-mover-leaver control, and no single place to answer "which agents can reach which data for whom". Enterprise-managed authorisation fixes this by making the corporate identity provider the decision point, using the Identity Assertion JWT Authorization Grant (ID-JAG) and, for MCP, the Enterprise-Managed Authorization extension. This guide is for identity and security teams deciding how agents should reach tools and data: it covers the model, the MCP extension, policy design, what the IdP can and cannot see, rollout and monitoring. The protocol detail is in the companion Identity Assertion JWT Authorization Grant Guide.

Key takeaways

  • The IdP becomes the approval point for agent access. The agent's client exchanges the user's SSO assertion for an ID-JAG only when IdP policy allows that client to reach that server for that user.
  • MCP has a stable extension for it. The Model Context Protocol extension io.modelcontextprotocol/enterprise-managed-authorization is marked stable and builds on ID-JAG draft -04, which as of 27 October 2026 is still an IETF Internet-Draft.
  • It is for agents acting for a present user. Okta's Cross App Access documentation excludes autonomous agents without a user session and machine-to-machine jobs; those need workload identity patterns.
  • Issuance is governed, usage is not. The IdP sees token issuance only, not the MCP traffic, so tool-level controls and logging still belong at the MCP server or gateway.
  • Revocation is the payoff. Because the resource side should not issue refresh tokens, disabling a user or changing policy at the IdP stops new access at the next token request.

The core MCP authorisation model is user-driven: each user authorises each MCP client to each MCP server. The MCP project's own description of Enterprise-Managed Authorization lists the problems that creates in companies: employees should not need to understand every server's authorisation details, security teams cannot enforce consistent policy when each user decides alone, onboarding means authorising dozens of services, and offboarding means revoking access service by service. See the MCP Authorisation Specification Guide for the baseline model.

Agents make each weakness worse:

  • Volume: one assistant may connect to many servers, and each connection is a stored grant.
  • Prompt-driven connection: an agent can ask a user to connect a new tool mid-task, and users approve consent prompts without reading scopes.
  • Concentration: agent platforms that hold refresh tokens for many users and services become high-value targets, the same pattern seen in the Salesloft Drift breach, where an integration vendor's stored OAuth access and refresh tokens were stolen and used to export customers' Salesforce data.
  • Invisible scope creep: users grant broad scopes once, and the agent keeps them indefinitely.

The OWASP agentic risk on identity and privilege abuse covers the same ground from the threat side. See the OWASP Agentic Top 10 ASI03 Guide.

How enterprise-managed authorisation works

  1. SSO to the agent's client. The user signs in to the MCP client (the AI assistant or agent host) through the enterprise IdP. The client keeps the ID Token or SAML assertion.
  2. Exchange for an ID-JAG. When an MCP server requires enterprise-managed authorisation, the client sends an RFC 8693 token exchange to the IdP requesting the ID-JAG token type. Per the MCP extension, audience must be the issuer of the MCP server's authorisation server and resource, if sent, must be the MCP server's RFC 9728 resource identifier.
  3. IdP evaluates policy. Group membership, role assignments and conditional access rules decide whether to issue. A denied user receives an error and the client never gets a token.
  4. Redeem at the MCP authorisation server. The client presents the ID-JAG with the JWT bearer grant. The server validates it (signature against the IdP's JWKS, audience, issuer, expiry, client match) and issues an access token restricted to the MCP server named in the grant.
  5. Call tools. The client calls the MCP server with that access token. The extension says the client must not redirect the user to the MCP authorisation server's authorisation endpoint.

Clients declare support for the extension through the MCP capabilities mechanism, and extensions are opt-in, never active by default. MCP publishes a client matrix showing which clients implement it; check it rather than assuming support.

ConcernPer-user consent (baseline MCP)Enterprise-managed (ID-JAG)
Who approves a client-to-server connectionEach userIdP administrator policy
Long-lived tokens held by the clientRefresh token per user per serverResource side should not issue refresh tokens; client holds its SSO session
OffboardingRevoke at each serverDisable at the IdP; next token request fails
Inventory of connectionsScattered across serversIdP issuance logs and policy
Scope controlWhatever the user acceptedIdP can narrow scopes; server can narrow further
Visibility of tool callsAt the server onlyStill at the server only

Designing the IdP policy for agents

Treat each agent client as an application in the IdP catalogue with its own access policy, and each MCP server or SaaS API as a resource it may be granted. Useful policy dimensions:

  • Client: which assistant or agent host is asking. Only approved, registered clients should be able to request ID-JAGs; the draft limits the flow to confidential clients.
  • Resource: which MCP server or API, identified by its authorisation server issuer and resource identifier.
  • User population: groups or roles allowed to use that client with that resource.
  • Scopes: the maximum tool or API scopes the IdP will put in the grant; keep write and admin scopes separate from read.
  • Authentication context: require a recent, phishing-resistant sign-in for sensitive resources. The IdP can respond with insufficient_user_authentication and a max_age to force step-up.
  • Conditions: device compliance, network or session risk where the IdP supports them.

Write these as a short register per resource: owner, data classification, allowed clients, allowed groups, maximum scopes, step-up requirement and review date. That register is what auditors and the AI governance board will ask for. See the Identity Governance for AI Agents Guide for review and certification.

Operating model and ownership

Enterprise-managed authorisation moves work from end users to named teams. Agree ownership before the pilot, or policies will be created once and never reviewed.

ActivityOwnerEvidence produced
Approve a new agent client for ID-JAGIdentity team with AI governance sign-offClient registration record, risk assessment, approved resources
Onboard a new MCP server or SaaS resourceResource owner (application or data owner)Resource register entry: classification, scopes, step-up rule
Grant a group access to a client and resource pairResource owner approves, identity team configuresIdP policy change ticket
Periodic review of policies and scopesResource owner, quarterly for sensitive dataCertification result in the IGA tool
Incident response for a misbehaving agentSecurity operationsIdP policy disabled for the client, timeline from IdP and server logs

The kill switch for a compromised or misbehaving agent client becomes a single IdP action: disable the client's policy, and no new ID-JAGs are issued. Existing access tokens at MCP servers stay valid until they expire, which is why access token lifetimes of minutes matter. See the AI Agent Kill Switch and Containment Guide.

What the IdP does not control

The MCP extension's security considerations are explicit that IdP visibility ends at token issuance and does not extend to traffic between client and server. That leaves gaps you must close elsewhere:

GapWhere to control it
Which tools the agent calls, with what argumentsMCP server authorisation per tool, or an MCP gateway. See the MCP Gateway and MCP Server Governance Guide
Prompt injection steering the agent to misuse a legitimate tokenTool-level least privilege, human approval for high-impact actions. See the Human-in-the-Loop for AI Agents Guide
Data returned to the modelServer-side filtering and data classification
Token theft from the clientShort access token lifetimes, sender-constrained tokens where supported
Agent acting with no user presentOut of scope for ID-JAG; use workload identity. See the Agent Identity and Access Guide

The ID-JAG draft allows an optional act claim but leaves actor semantics to future profiles. Until a profile defines how the agent itself is identified in the grant, do not build authorisation decisions that depend on distinguishing "the user" from "the agent acting for the user" inside the ID-JAG; identify the agent through its client ID and through server-side logging instead.

Account linking at the MCP server

The MCP guidance is to use the ID-JAG's subject claim as the stable primary identifier and to fall back to the email claim only to match accounts created before enterprise-managed authorisation was configured. In practice:

  • Record the pair (issuer, subject) on first successful link and use it from then on.
  • Use email once, for migration, and only when the IdP is authoritative for that domain.
  • Handle multi-tenant IdPs with the tenant claim so the same subject value in two tenants never maps to one account.
  • Provision ahead with SCIM where possible so the server knows the user before the first agent call.

Choosing the right pattern for an agent

Agent situationPattern
Assistant used interactively by an employee, calling corporate MCP servers or SaaSEnterprise-managed authorisation with ID-JAG
Consumer or personal assistant, no enterprise IdPStandard MCP authorisation with user consent
Scheduled or autonomous agent with no user sessionWorkload identity and client credentials, with its own NHI owner. See the Machine-to-Machine OAuth (Client Credentials) Guide
Agent calling services inside your own platform after authorisationTransaction Tokens for the internal call chain. See the OAuth Transaction Tokens Guide
Agent crossing to a partner domain with its own authorisation server and no shared IdPIdentity chaining. See the OAuth Identity and Authorization Chaining Across Domains Guide

Small organisations should start by restricting which agent clients may be used at all and enabling enterprise-managed authorisation where their IdP and chosen assistant already support it. Large organisations should make support a procurement requirement for agent platforms and internal MCP servers, and run the resource register as part of the AI governance process.

How to implement

  1. Inventory agent connections. List agent clients in use and the MCP servers and SaaS APIs they reach, including consents already granted by users.
  2. Check support. Confirm your IdP issues ID-JAGs, your agent clients support the MCP extension (client matrix), and your MCP servers' authorisation servers advertise urn:ietf:params:oauth:grant-profile:id-jag in authorization_grant_profiles_supported.
  3. Pick one internal MCP server with read-only tools and a pilot group.
  4. Configure policy at the IdP, register the client at the MCP authorisation server, and set access token lifetime to minutes with no refresh tokens.
  5. Test denials: a user outside the group, a disabled user, an unapproved client, a scope above policy, an ID-JAG for a different server. All must fail.
  6. Pilot and compare against the old consent-based connection, then revoke those legacy grants for pilot users.
  7. Expand server by server, adding write tools last and with step-up or human approval.
  8. Rollback: re-enabling user consent for the client is the fallback; keep it disabled by default and time-boxed if used.

Detection and measurement

  • IdP: ID-JAG issuance and denials by client, resource, user and policy rule.
  • MCP authorisation server: grants redeemed, account links created, email-fallback links, and any refresh token issued.
  • MCP server or gateway: tool calls with user subject, client ID and token identifier, so a tool call can be traced to the ID-JAG that authorised it.
  • Alerts: ID-JAG requests for unapproved resources, sudden growth in resources per user, redemptions after IdP deprovisioning, and grants from clients outside the approved list.

Metrics: percentage of agent-to-tool connections governed by IdP policy; number of user-consented grants remaining for managed clients; time from deprovisioning to last successful tool call; ID-JAG denial rate by reason; resources per agent client against the approved register.

Framework mapping

FrameworkReferenceRelevance
NIST SP 800-53 Rev 5AC-2 Account Management; AC-3 Access Enforcement; AC-6 Least Privilege; AU-12 Audit Record GenerationCentrally approved, least-privilege agent access with traceable records
NIST CSF 2.0PR.AA-01, PR.AA-05; DE.CM-03Managed identities, policy-defined access, monitoring of technology usage
ISO/IEC 27001:2022 Annex A5.15 Access control; 5.18 Access rights; 8.15 LoggingAccess rights lifecycle and evidence
CIS Controls v8.1Control 6 Access Control Management; Control 8 Audit Log ManagementCentral access control and logs
MITRE ATT&CKT1528 Steal Application Access Token; T1550.001 Application Access TokenReduced store of long-lived agent tokens

Common failure patterns

  • Leaving per-user consent available for managed clients, so users and agents bypass IdP policy.
  • Treating issuance control as runtime control. An approved token can still be used for a harmful tool call after prompt injection; tool-level policy is still required.
  • Using ID-JAG for headless agents by keeping a service user permanently signed in, which recreates a shared credential with a human label.
  • Broad scopes in the first policy that are never narrowed after the pilot.
  • Email-only account linking on MCP servers, enabling takeover where addresses are reused.
  • Long access token lifetimes at MCP servers. If the MCP authorisation server issues hour-long or day-long tokens, disabling a user or client at the IdP takes effect only when those tokens expire, and the kill switch is slow exactly when it is needed. Keep lifetimes short and use token revocation where the server supports it. See the OAuth Token Introspection and Revocation Guide.
  • Assuming client support without checking the client matrix and the draft version each component implements.

Practitioner checklist

  • Inventory agent clients, MCP servers and existing user consents.
  • Require enterprise-managed authorisation support in agent platform procurement.
  • Keep a resource register: owner, classification, allowed clients, groups, scopes, step-up.
  • Configure IdP policy per client and resource; start read-only.
  • Issue short-lived access tokens with no refresh tokens at MCP authorisation servers.
  • Link accounts on issuer plus subject; use email only for migration.
  • Disable user consent for managed clients and revoke legacy grants.
  • Log tool calls with subject and client ID and correlate with IdP issuance.
  • Use workload identity, not ID-JAG, for agents without a user session.
  • Track the ID-JAG draft and MCP extension versions in use.

Standards and references

Related NHI Mgmt Group resources: Identity Assertion JWT Authorization Grant Guide · MCP Authorisation Specification Guide · AI Agent Authorisation Guide · OAuth Token Exchange and Delegation Guide

PDF for subscribers

Take this guide with you

Every section, table and checklist in one PDF, with links to the sources and related guides. Subscribe to our weekly NHI & AI Identity Journal and we’ll email you a download link.

Free weekly newsletter

Subscribe to the NHI & AI Identity Journal

The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.

Bonus 33% off our NHI Course when you subscribe.

NHIMG Editorial Note
Researched and written by the NHI Mgmt Group editorial team using AI tools, under the editorial direction of Lalit Choda. Facts are drawn from the sources listed on this page and checked by our automated quality process. Last updated 27 October 2026. Spotted an error? Use our contact page.
NHI Mgmt Group, the independent authority on Non-Human Identity, IAM and Agentic AI security. nhimg.org

PDF Downloads for Newsletter Subscribers

Get the guide

Enter your details and we’ll email you a download link.

The full guide is free to read online. The PDF is a thank you for subscribing.

Fill in all fields marked * and tick the PDF box to continue.