Yes, you could define A RAR type with all of the parameter fields you want:

{
  type: example.com/agent-ai,
  provider: ...
  model: ...
  capability: ...
}

And whatever other fields you'd want.

I can't say this strongly enough, don't use scopes in the way you're proposing.

- Justin
________________________________
From: Hemanth H.M <[email protected]>
Sent: Thursday, December 25, 2025 4:36 PM
To: Justin Richer <[email protected]>
Cc: [email protected] <[email protected]>
Subject: Re: [OAUTH-WG] [New I-D] draft-hemanth-oauth-ai-scopes-00 - OAuth 2.0 
Extension for AI Model Access

Hey Warren,

Good question. Current OAuth doesn't have a standard way to scope access *to 
specific models* or attach usage limits (spend/rate) directly to the token 
metadata without heavy custom extensions, right? This ID tries to standardize 
that delegation layer.

Justin, We can leverage RAR type for this?


--
Thank you,
Hemanth.HM<http://www.h3manth.com>



On Thu, Dec 25, 2025 at 1:31 PM Justin Richer 
<[email protected]<mailto:[email protected]>> wrote:
It is an extremely terrible idea to create a structure for scopes. I've done 
this several times in different ecosystems and it always starts out ok but 
falls apart quickly. Do not repeat this mistake.

If you need structure for access, define a RAR type, that's what it's there for.

- Justin
________________________________
From: Hemanth H.M <[email protected]<mailto:[email protected]>>
Sent: Wednesday, December 24, 2025 4:41 PM
To: [email protected]<mailto:[email protected]> 
<[email protected]<mailto:[email protected]>>
Subject: [OAUTH-WG] [New I-D] draft-hemanth-oauth-ai-scopes-00 - OAuth 2.0 
Extension for AI Model Access

Hi OAuth WG,

I've submitted a new Internet-Draft for your consideration:

draft-hemanth-oauth-ai-scopes-00 - OAuth 2.0 Extension for AI Model Access

Problem: AI model APIs (OpenAI, Anthropic, Google, etc.) require API key 
delegation, but current practices involve sharing master keys directly with 
third-party applications—no scoping, no revocation, no usage limits.

Proposal: Extend OAuth 2.0 with:


  1.  Standard scope syntax: ai:<provider>:<model>:<capability>
  2.  Token metadata for spend/rate limits
  3.  Token introspection extensions for usage tracking
  4.  Security considerations (DPoP/mTLS for high-security deployments)

GitHub: https://github.com/hemanth/oauth-ai-scopes

I'd welcome feedback on the scope syntax, alignment with existing OAuth 
extensions (RFC 8707, RFC 9449), and whether this is something the WG would 
consider adopting.

P.S: I also started https://okap.dev as a separate protocol, in case...

--
Thank you,
Hemanth.HM<http://www.h3manth.com>

_______________________________________________
OAuth mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to