Review · November 22, 2026
A Discovery Document and Signed Catalogue for Capability Marketplaces Summary
The Problem
Agent and tool indexes are run by many independent operators. Each one publishes its basic facts at a different well-known path with a different schema. Those that sign their catalogues rarely say which bytes the signature covers.
This draft defines one JSON discovery document for an index. It carries the index's name, protocol versions, catalogue URL, and the public key that signs the catalogue. It also defines one signature format for that catalogue and tells readers to pin the key.
How It Fits In
Half of this draft is in the DAWN charter and half is not. The draft says so itself in Section 1.2.
The charter allows discovery of "a service such as a catalog that will further the discovery process". The discovery document fits that. The charter excludes "catalogs of capabilities" in the initial phase. The catalogue format, the signature profile, and key pinning fall on that side. The author calls them "adjacent work" and keeps them in because a reader needs them to authenticate the catalogue.
It is the first half of a two-draft set. The author's index federation draft builds on it for crawling between indexes. This draft in turn sends readers to the federation draft for its SSRF rules, which is a normative requirement pointing at an informative reference.
It builds on Well-Known URIs (RFC 8615), HTTP, JSON, and Ed25519 (RFC 8032). There is no OAuth or WIMSE dependency. It does not use JOSE, COSE, or HTTP Message Signatures.
It connects to other DAWN drafts in two places. A DNS-AID _index._agents SVCB record can point to the discovery document. An appendix maps the document onto the MDI model.
It registers one well-known URI, agent-marketplace. It adds no new roles beyond "index" and "reader".
High-Level Summary
Discovery document. An index serves a JSON object at /.well-known/agent-marketplace. Three members are required: name, protocol_versions, and catalogue_url. If the catalogue is signed, signer_public_key and signature_profile are required too. An optional peers list points at other indexes.
Catalogue. The catalogue is a JSON object at catalogue_url, possibly on a different host. The draft sets a minimum: a protocol version, a timestamp, an array of capabilities with a name, description, and input schema, and a signature object. Everything else depends on the protocol version, which this draft does not define.
Signature profile. The one defined profile is ed25519-pipe-2. It builds a pipe-separated string from five values: a count, a timestamp, a version, and SHA-256 digests of two JSON values. The digests use a custom JSON serialisation. The Ed25519 signature covers that string. Verifiers must reject small-order keys and non-canonical signatures before checking.
Key pinning. A reader pins the key from the discovery document on first contact. After that, a different key is never accepted automatically. A human operator has to approve it.
So what does a verified signature tell you? Only that the same key holder published the catalogue as last time. The draft is clear that it says nothing about who the operator is, or whether the listed capabilities exist or cost what they claim.
Significant Interoperability Concerns
The protocol versions point at nothing. protocol_versions are opaque strings. The example uses "v1" and "v2". No document defines what those mean, and there is no registry. A reader that recognises none of them must not use the index, so every reader written from this draft alone must walk away from every index.
The signature covers a subset of the catalogue. Only five values go into the signed string. Any other top-level member of the catalogue is unsigned. The draft says the full schema varies by protocol version, so each version's extra members can be changed in transit without breaking the signature.
The signed string references things the draft does not define. by_hub is described only as "per-peer metadata in the deployed format". capabilities_count is signed but is not in the catalogue's minimum content. The discovery document has a member with the same name, and the draft says a mismatch there must not be treated as an error. An implementer has to guess which one is signed.
The JSON serialisation is custom and incomplete. It is Python's default json.dumps output with sorted keys. The draft acknowledges it differs from JCS (RFC 8785) and keeps it to preserve existing signatures. String escaping is not specified. Array separators are not specified. Number formatting says "shortest decimal form that round-trips", and Python and JavaScript disagree on exponent notation for that. Two honest implementations will produce different digests for the same catalogue.
The small-order point list is not implementable from the text. The encodings are abbreviated, such as 26e8958f..fc05. The full values live in a GitHub repository of test vectors, which is an informative reference.
Key rotation breaks every reader. A new key needs a manual decision at each reader. The suggested fix is dual signing "using additional members of the signature object", and those members are not defined. Many readers here are autonomous agents with no operator watching.
Stripping the signature is not clearly handled. A catalogue whose signature is not an object "MUST be treated as unsigned". The draft never says what a reader with a pinned key does with an unsigned catalogue. Signing is also optional overall, while the DAWN charter requires integrity for all discovery information.
There is no replay protection. There is no validity window or expiry. The draft notes a reader "can" compare generated_at with the last catalogue it saw, and requires nothing.
The DNS example is thin. It is one SVCB record with no statement of what a reader does if the well-known suffix is something else.
Mechanisms That Could Be Removed
The custom signature profile. JWS over the exact catalogue bytes needs no canonicalisation at all. The whole of Section 5 exists to work around a problem that signing the bytes avoids. HTTP Message Signatures with a Content-Digest is another existing option. Either one also brings key identifiers and algorithm agility, and a JWK Set URL in place of a single raw key would make rotation a non-event.
capabilities_count and generated_at in the discovery document. Both are informative, and readers are told to ignore mismatches. They carry nothing a reader can act on.
peers. This draft "gives them no further meaning". It belongs in the federation draft.
User interface rules. Two MUST NOTs govern what a reader's UI may suggest about a signature. They are good advice and are not testable protocol behavior.
The HEAD requirement. Nothing in the draft uses it.
The catalogue itself. The discovery document is the in-charter piece. Splitting it from the catalogue and signature work would give DAWN a three-member JSON document it could actually evaluate. As written, the in-scope part is about a page and the out-of-scope part is the rest.
Writing Quality
The writing is good. It is 18 pages, uses BCP 14 keywords correctly, and reads like an IETF document. Each requirement gets a sentence of justification. The "Implementation experience" appendix ties four requirements to real failures, which is more than most drafts offer.
The author is candid. The draft states that no other party is known to have implemented from the text, that pinning does not identify anyone, and that rotation is unsolved.
The naming is inconsistent. The title says "capability marketplaces", the text says "index", the well-known URI says "agent-marketplace", and the signed string says "tools" and "hub". The spelling "catalogue" differs from the charter's "catalog".
The compromises all point the same way. The serialisation rule, the pair names in the signed string, and the by_hub digest are kept because the deployed system already does them. This is a description of one implementation, lightly renamed.
The Implementation Status section reports six public indexes, all run by the author or affiliates, plus three more implementations by the same author. Two unrelated services publish catalogues that verify, running unknown software.
The draft has had four revisions in about a week. The last two added charter and MDI positioning without changing any requirement.
Background of the Author
Aleksandr Artamokhov (Independent) has no RFCs. They have two active Internet-Drafts, this one and its companion federation draft, and no expired drafts. They are a new IETF author and are actively iterating on both.
Bottom Line
The discovery document is small, reasonable, and arguably in scope for DAWN as a description of a catalog service. It overlaps with what DNS-AID's index record and the ARD drafts are already doing, so the working group will need to pick one shape for "here is an index".
The signature profile is the weak part. It standardises one deployment's wire format, including a hand-rolled JSON canonicalisation, where JWS would do the job with existing libraries. The useful feedback to the author is to split the draft, keep the discovery document, and replace the signature profile with an existing IETF signature format.