Max,
This is a useful decomposition, and I think your own parenthetical identifies
the crux: the workaround makes the gateway the subject, and everything upstream
of it sees only the gateway. That is not an implementation wrinkle, it is the
general behavior of re-subjecting intermediaries: each mediation hop that
acquires authority under its own identity collapses the chain behind it, and
whatever the deployment needed the chain for (per-agent policy, audit
attribution, differential revocation) has to be reconstructed some other way.
The actor_token idea does help, but it is worth being precise about what it
buys and what it costs. It records one hop: agent plus gateway, bound at mint
time. Generalized to deeper topologies (agent behind agent behind gateway,
which is the ordinary MCP composition case), it becomes stacked act semantics
minted by the IdP per hop, which means the IdP is the recorder of the chain and
every extension of the chain is a round-trip to it. That is a coherent
architecture, and for the fullest articulation of it I would point at Karl's
Identity Continuation Assertion note, which works out the caller-pushed,
control-plane-resident version of exactly this direction, including its
revocation story. The trade is then explicit: chain state lives at the control
plane, per-agent visibility is restored to the IdP, and the chain's length is
priced in round-trips.
One assumption in the scenario deserves flagging, because it is where the hard
version of the problem lives: the construction works because the agent, the
gateway, and the upstream all sit in one IdP's orbit, so the actor_token is
validatable by its own issuer. When the gateway and the agent belong to
different organizations (the multi-vendor gateway case, which the DMSC BoF
decks this week treated as the normal deployment), validating that actor_token
is itself the cross-domain trust-establishment question, and we are back at the
cold-start gap the use-cases draft now names in UC5.
Which is my main suggestion: this thread has produced better deployment
evidence for that gap than anything currently in the draft. Meiling has opened
issues #17 (UC5 alignment) and #18 (the cross-organizational companion case) on
the use-cases repository; the gateway-collapse observation and the one-orbit
assumption both belong there as recorded evidence, and I would encourage you to
add them. The mechanism questions will sort themselves out per draft; the
requirements record is what seems to be collecting this year
Morgan
On Thursday, July 23rd, 2026 at 12:34 AM, Max Gerber
<[email protected]> wrote:
> Thinking out loud: One way to do this today is for the MCP Gateway to also
> act as a client to the IDP, requiring the end user to log in to both the
> Gateway and the Agent. This can be done with URL Elicitation in the MCP
> space, or if the Gateway runs its own OAuth AS as a shim. The Agent gets a
> regular access token, and the Gateway gets an ID Token with the Gateway
> client as the audience.
>
> When the Agent makes requests to the Gateway, the Gateway can use the ID
> Token to retrieve an ID-JAG for the upstream MCP AS. This ID-JAG will only
> identify the Gateway client, not the Agent itself. This would make the
> Gateway responsible for access control decisions & auditing, and it cuts off
> the IDP from observing per-Agent behavior - which undoes a lot of the
> benefits of ID-JAG in the first place.
>
> Would having the Gateway pass in the Agent's access token as an `actor_token`
> in the token exchange request to the IDP help here? That would give the IDP a
> mechanism to identify the Agent (and also bind both the Agent identity +
> Gateway identity within the ID-JAG). Actor tokens are deliberately out of
> scope of the ID-JAG spec, but they could be defined in a future profile.
>
> On Thu, Jul 23, 2026 at 3:58 AM morganLR
> <[email protected]> wrote:
>
>> Eric,
>>
>> Good question, it is my understanding that the two paths you named do
>> different jobs, so I'd suggest both, in this order given the timing this
>> week.
>>
>> 1. GitHub issue against the ID-JAG repo, today if you can. That's the right
>> vehicle for the narrow question: does ID-JAG address this topology, or is it
>> out of scope by design?
>>
>> 2. The use-case record, for the broader point. The chairs have been clear
>> that use cases and requirements come before mechanism selection, and there
>> is an active collection effort (draft-chen-oauth-agent-authz-use-cases is
>> the current gathering point).
>>
>> 3. If you want it durable under your own name, a short individual
>> problem-statement I-D is always an option.
>>
>> Happy to look over the issue text before you file it if that's useful.
>>
>> Best,Morgan
>> On Wednesday, July 22nd, 2026 at 2:12 PM, Eric Leleu
>> <[email protected]> wrote:
>>
>>> Morgan, Karl,
>>>
>>> Thank you both for the detailed and thoughtful responses. I'll block out
>>> time to properly go through the references you both pointed to.
>>>
>>> Morgan, one practical question on process: you mentioned this topology
>>> deserves to be in the record the working group is building this week.
>>> I'm not familiar with the actual procedure for that. Should I open a GitHub
>>> issue against the ID-JAG repo, or write it up as a separate
>>> use-case/problem statement?
>>> If the latter, where should that go? Happy to follow whichever path is
>>> expected.
>>>
>>> Thanks again,
>>> Eric
>>>
>>> Le mer. 22 juil. 2026 à 18:44, Karl McGuinness <[email protected]> a
>>> écrit :
>>>
>>>> Extending ID-JAG to support public clients is an open topic for
>>>> discussion. I encourage you to read these issues and chime in with
>>>> additional use cases, concerns, or feedback
>>>>
>>>> https://github.com/oauth-wg/oauth-identity-assertion-authz-grant/issues/85
>>>> https://github.com/oauth-wg/oauth-identity-assertion-authz-grant/issues/113
>>>>
>>>> The gateway scenario you outlined is similar to one where folks need to
>>>> make a second hop after exchanging the ID-JAG for an access token and no
>>>> longer have access to the original identity assertion to obtain a new
>>>> ID-JAG. I've been exploring this scenario as an identity continuation
>>>> pattern
>>>> (https://notes.karlmcguinness.com/notes/identity-continuation-assertion/).
>>>> A key premise of ID-JAG is that re-subjecting across trust domains is a
>>>> mint and not an attenuation
>>>> (https://notes.karlmcguinness.com/notes/re-subjecting-is-a-mint-not-an-attenuation/).
>>>> This may not be relevant to your specific gateway deployment scenarios,
>>>> but it is common in deployments where a gateway fronts existing SaaS apps
>>>> for example.
>>>>
>>>> I'm not super happy with the solution I arrived at
>>>> (https://mcguinness.github.io/draft-mcguinness-oauth-id-continuation-assertion/draft-mcguinness-oauth-id-continuation-assertion.html
>>>> and for this reason I haven't submitted it yet as an I-D because it still
>>>> needs workshopping. I'm sharing this only to see if it aligns with your
>>>> use case. I expect smarter people will find better solutions as this
>>>> particular solution isn't pretty.
>>>>
>>>> For scenarios that don't need re-subjecting then an attentuation based
>>>> soluton may be attractive but that it outside the scope of the ID-JAG
>>>> profile.
>>>>
>>>> -Karl
>>>>
>>>> On Wed, Jul 22, 2026 at 8:58 AM morganLR
>>>> <[email protected]> wrote:
>>>>
>>>>> Eric,
>>>>>
>>>>> The dilemma is real and I do not think you missed a mechanism. It is
>>>>> worth naming why it arises structurally: an identity assertion is
>>>>> audience-locked to the party it was issued to, and delegation through an
>>>>> intermediary requires either re-minting at the issuer or a way for the
>>>>> intermediary's participation to be first-class in the exchange. Section
>>>>> 4.3.3 as written admits neither, so every gateway hop reproduces your
>>>>> dilemma.
>>>>>
>>>>> Two resolution families exist, and the draft could name which it intends.
>>>>>
>>>>> Within the current model, RFC 8693 already has the missing piece: the
>>>>> actor_token. A composite exchange where the gateway authenticates as
>>>>> itself, presents the agent's assertion as subject_token and its own
>>>>> credential as actor_token, and the IdP applies the audience check to the
>>>>> subject token's original audience while separately authorizing the actor
>>>>> as a registered intermediary for that audience, resolves your topology
>>>>> without new machinery. The cost is that the IdP must now hold policy
>>>>> about which intermediaries may act for which clients, and the exchange
>>>>> remains a per-upstream, per-hop round trip to the IdP: the gateway mints
>>>>> one ID-JAG per upstream AS, on the critical path.
>>>>>
>>>>> The other family carries the delegation as a verifiable artifact instead
>>>>> of re-minting it: the agent delegates narrowed authority to the gateway
>>>>> in a form the upstream can verify directly against the original issuer's
>>>>> key, so the intermediary attenuates rather than exchanges, and no per-hop
>>>>> trip to the IdP exists. That property class (attenuation verifiable from
>>>>> the artifact alone, no runtime callback to the issuing side, possession
>>>>> bound at each hop) is what R1, R3, and R4 in
>>>>> draft-reece-wimse-cross-org-delegation describe, and gateway aggregation
>>>>> is among the cleanest motivating topologies for it.
>>>>>
>>>>> Either way, your topology deserves to be in the record the working group
>>>>> is building this week: the chairs have asked for use cases and problem
>>>>> statements before solutions, and agent-behind-gateway with a
>>>>> public-client agent is arguably the most common deployment shape MCP has
>>>>> produced. I would encourage writing it up exactly as you have here.
>>>>>
>>>>> Morgan
>>>>>
>>>>> On Wednesday, July 22nd, 2026 at 7:23 AM, Eric Leleu
>>>>> <[email protected]> wrote:
>>>>>
>>>>>> Hi all,
>>>>>>
>>>>>> I'd like to raise a topology question about the ID-JAG draft
>>>>>> (draft-ietf-oauth-identity-assertion-authz-grant) that I expect to come
>>>>>> up often as agentic/MCP deployments grow, and I don't think the current
>>>>>> text resolves it from my understanding.
>>>>>>
>>>>>> Topology Description:
>>>>>> ----------------------------
>>>>>>
>>>>>> - An AI coding agent (e.g. Claude Code) acts as an MCP Client. It is
>>>>>> registered at the enterprise IdP as a *public* client (native/CLI app,
>>>>>> PKCE, no client secret).
>>>>>> - The agent authenticates directly against the enterprise IdP. As per
>>>>>> OIDC Core, the resulting ID Token's `aud` is the agent's own client_id.
>>>>>> In the same flow the agent also obtains an access token whose
>>>>>> audience/resource is an MCP Gateway sitting in front of it.
>>>>>> - The MCP Gateway aggregates several upstream MCP servers. Each upstream
>>>>>> server is protected by its own Resource Authorization Server, and each
>>>>>> of those independently trusts the same enterprise IdP.
>>>>>> - Per ID-JAG, to reach a given upstream server the Gateway needs an
>>>>>> access token minted through the ID-JAG flow: Token Exchange at the IdP
>>>>>> to mint an ID-JAG (aud = upstream AS) and then execute JWT-Bearer
>>>>>> redemption at that AS.
>>>>>>
>>>>>> The conflict :
>>>>>> -----------------
>>>>>>
>>>>>> Section 4.3.3 [1], "IdP Token Exchange Processing Rules", requires that
>>>>>> when the subject_token presented in the Token Exchange call is an
>>>>>> identity assertion (ID Token), "the IdP MUST validate the assertion and
>>>>>> MUST validate that the audience of the assertion (e.g. the aud claim of
>>>>>> the ID Token or SAML Audience) matches the client_id of the client
>>>>>> authentication of the request". In the Gateway topology above, this rule
>>>>>> creates a dilemma where neither party can execute the exchange:
>>>>>>
>>>>>> 1. The agent is the audience of the ID Token, but the agent has no
>>>>>> reason to call Token Exchange itself — it doesn't know which upstream
>>>>>> AS/audience a given MCP tool call needs, and as a public client it's
>>>>>> also not well positioned to be trusted with that capability directly. As
>>>>>> defined in the draft "this specification SHOULD only be supported for
>>>>>> confidential clients."
>>>>>> 2. The Gateway is the party that actually needs the ID-JAG (it knows the
>>>>>> target upstream AS), and it is the Gateway that would authenticate the
>>>>>> Token Exchange call. But the ID Token's `aud` is the agent's client_id,
>>>>>> not the Gateway's. Per the mandatory audience check above, the IdP must
>>>>>> reject that exchange no matter how the ID Token reaches the Gateway
>>>>>> (forwarded by the agent or otherwise).
>>>>>>
>>>>>> From my understanding, neither party in this agent-to-gateway
>>>>>> architecture can legally execute the ID-JAG exchange under the current
>>>>>> specification.
>>>>>>
>>>>>> Did I miss a mechanism already defined in the draft (or in
>>>>>> draft-ietf-oauth-identity-chaining) to handle this delegation/proxy
>>>>>> pattern?
>>>>>> Is there ongoing work to address multi-hop or API Gateway patterns where
>>>>>> the identity assertion's audience differs from the calling proxy's
>>>>>> client_id?
>>>>>>
>>>>>> Thanks,
>>>>>> Eric
>>>>>>
>>>>>> [1]
>>>>>> https://datatracker.ietf.org/doc/html/draft-ietf-oauth-identity-assertion-authz-grant#name-processing-rules
>>>>>> --
>>>>>>
>>>>>> Eric LELEU
>>>>>>
>>>>>> Staff Software Engineer
>>>>>>
>>>>>> E [[email protected]](mailto:[email protected])
>>>>>>
>>>>>> Hold Nothing Back
>>>>>>
>>>>>> http://youtube.com/c/Graviteesource?sub_confirmation=1https://www.linkedin.com/company/gravitee-io
>>>>>
>>>>> _______________________________________________
>>>>> OAuth mailing list -- [email protected]
>>>>> To unsubscribe send an email to [email protected]
>>>
>>> --
>>>
>>> Eric LELEU
>>>
>>> Staff Software Engineer / AM Teach Lead
>>>
>>> E [[email protected]](mailto:[email protected])
>>>
>>> Hold Nothing Back
>>>
>>> http://youtube.com/c/Graviteesource?sub_confirmation=1https://www.linkedin.com/company/gravitee-io
>>
>> _______________________________________________
>> OAuth mailing list -- [email protected]
>> To unsubscribe send an email to [email protected]
_______________________________________________
OAuth mailing list -- [email protected]
To unsubscribe send an email to [email protected]
