Discovery of Model Context Protocol Servers via DNS TXT Records

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶

This Internet-Draft will expire on 23 January 2027.¶

Copyright Notice

Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶

This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document.¶

Table of Contents

1. Introduction

Model Context Protocol (MCP) [MCP] is an open protocol for structured interaction between AI agents and tool-providing servers. A complete agent-to-organisation-to-individual interaction chain has three distinct discovery requirements:¶

1. Service discovery. Where is the MCP server endpoint? What transport does it speak? What cryptographic key authenticates it? This is the question v01 of this document answers via the _mcp.<domain> record.¶

2. Organisational identity bootstrap. Who is the organisation operating the server? What is its legal entity? Where is it registered? Under what regulatory frameworks does it operate, and which automated access pathways must it refuse to participate in? This is the question v02 answers via the _org-alter.<domain> record.¶

3. Individual identity recognition. Who is the Sovereign-tier person bound to a ~handle hosted under the domain? What public key signs their statements? What append-only log anchors the lifecycle of their identity? How may their envelope be revoked? This is the question v03 introduces via the _alter.<domain> record.¶

The three questions are distinct. An MCP client may need to discover an endpoint without caring about the operator's identity or any individual handle. An onboarding wizard installing an org-alter instance may need to read the operator's identity without caring (yet) about the MCP endpoint. A recognition verifier (resolving an alter:~alice URI) needs the individual envelope without necessarily invoking an MCP session. Conflating any two of these into a single TXT record would force every consumer to parse fields it does not need and would crowd the 255-octet character-string limit. Splitting them across three underscore-prefixed labels mirrors the pattern established by DKIM [RFC6376] (_domainkey._domain) and DMARC [RFC9989] (_dmarc._domain), where each record serves a single semantic purpose. SPF [RFC7208] and MTA-STS [RFC8461] set the same precedent for carrying policy in a TXT record under a reserved label.¶

This revision is fully backward-compatible with v01 and v02. Implementations that consume only the _mcp.<domain> record continue to work unchanged. Implementations that wish to bootstrap an org-alter identity may additionally query _org-alter.<domain>. Implementations that wish to recognise an individual ~handle may additionally query _alter.<domain>.¶

This document stands alone. Earlier revisions deferred the envelope schema, the JSON Canonicalisation Scheme (JCS, [RFC8785]) serialisation, and the verification algorithm to a companion specification a reader could not fetch, which made this document unimplementable by anyone outside its author's organisation. They also incorporated the _mcp and _org-alter record formats and procedures by reference to revisions 01 and 02, by section numbers that named the wrong sections, so a reader who followed a pointer arrived at the wrong place in the right document. Both deferrals are removed.¶

All three record formats (Section 5, Section 6, Section 7), all three procedures (Section 12.1, Section 12.2, Section 12.3), the signing input (Section 7.4) and the caching rules (Section 13) are given here in full. An implementer needs no other document in this series. Revisions 01 and 02 are cited where this document records what they said, and are cited informatively, because nothing here depends on fetching them.¶

The individual-identity layer is grounded, as the organisational layer is, in the identity field framework of [MORRISON-IFT]. A ~handle is not a reserved alphanumeric slot but a durable recognition attractor in the identity field. A DNS record provides a discrete checkpoint into that field: the envelope published at _alter.<zone> is the handle-holder's own canonical declaration, signed by their Ed25519 key and consumable by any resolver with access to a DNSSEC-validating recursive resolver. It also carries a reference to an IdentityLog root, whose status as a trust anchor Section 10 sets out honestly and narrowly.¶

2. Terminology

(Terminology from v01 and v02 is retained. Additional terms introduced in this revision are defined below.)¶

Envelope

An Ed25519-signed JSON object binding a ~handle to a public key, an IdentityLog root reference, an inception timestamp, a revocation hash commitment, a signature algorithm tag, a detached signature, and a caveats array. The envelope is the unification primitive of the ALTER identity architecture. The TXT grammar that carries the seven required envelope fields across DNS, and the signing input over which they are verified, are specified in Section 7 of this document.¶

~handle

A Sovereign-tier identifier, leading tilde mandatory (e.g. ~alice). Bot-tier handles carry a .bot suffix (e.g. ~example-bot.bot); Instrument-tier handles use the prefix ~cc- (e.g. ~cc-example-model).¶

IdentityLog

An append-only transparency log recording envelope lifecycle events (mint, caveat-add, revocation, key-rotation), whose root the ilr= field of Section 7 references. Revision 04 of this document described a Signed Tree Head emitted every minute and cross-anchored to four named witness surfaces. No such witness federation exists, and the description is withdrawn. The log's cadence, its witness protocol, and the client profile for checking it are NOT specified by this document and are not yet specified anywhere a reader can fetch. Section 10 states what a resolver may and may not conclude from ilr= in the meantime.¶

Organ

A broadcast surface for the envelope. The three organs are: DNS publication (this document), the local alter-runtime L3 daemon, and a hardware-anchored device-organ quorum. The term is canonical; do not substitute "channel", "vector", or "emitter" at the architectural level.¶

Recognition

The act of a resolver observing and verifying an envelope on cryptographic merit. Recognition is distinct from claim: publishing a TXT record is not a claim of identity, it is an observable assertion that the resolver may verify or reject. No field of this document carries a claim verb; resolvers recognise envelopes, they do not honour publisher assertions about them.¶

DNSSEC Validation

The act of an authenticating DNS resolver verifying the RRSIG chain from the root trust anchor to the TXT RRset, per [RFC4033], [RFC4034], and [RFC4035], and setting the AD bit on the response delivered to the stub client.¶

DANE TLSA Pin

A DNS TLSA resource record [RFC6698] binding a server's leaf TLS certificate (or the public key therein) to the zone that hosts the envelope. In this document, the pin applies to the MCP endpoint at mcp.<zone>.¶

alter: URI

A dispatch URI scheme provisionally registered with IANA per [RFC7595]. Handler guidance is given in Section 11 of this document.¶

(Terms from v02, Org-Identity Record, Identity Bootstrap, Canonical Entity Identifier, Regulatory Refusal Marker, are retained.)¶

3. Related Work

Several drafts now publish agent discovery data in DNS, and they arrived at overlapping vocabularies independently.¶

Two of them are called AID and they are different documents. [AGENT-AID] is "Agent Identity and Discovery", and [DNS-AID] is "DNS for AI Discovery". A reader who meets both should not have to work that out from context, so this document names them apart.¶

[AGENT-AID] is the closest neighbour to the service-discovery record of this document, and it came first. It publishes a TXT record at _agent.<domain> carrying a leading v=aid1 version tag and semicolon-delimited key=value pairs, which give the endpoint URI (u), a protocol token (p), an authentication hint (a), and optionally an Ed25519 public key for endpoint proof (k) with a rotation identifier (i). The _mcp record of Section 5 reaches the same three facts, an endpoint, a protocol, and a key, in an underscore-prefixed TXT record with a leading version tag and semicolon-delimited pairs, and both patterns are derived from DKIM [RFC6376]. Neither document copied the other. Two drafts landed on one shape because DKIM is where anyone publishing a key in TXT learns to look, and [AGENT-AID] published that shape first, in March 2026.¶

The two compose, and [AGENT-AID] says how. Its abstract states that AID is intentionally small and that after discovery, protocol-specific mechanisms such as MCP handle communication and capability negotiation, and its protocol registry already carries mcp as a token. This document is what that token defers to. [AGENT-AID] answers where an agent is and which protocol to speak to it; the _mcp record answers what an MCP endpoint then requires, its transport, its authentication, its capability scope and its key epoch; and the _org-alter record of Section 6 answers who is behind the endpoint, with a legal entity, a registry identifier, operating regions and regulatory refusal markers. A deployment may publish both, and a client that learns p=mcp from _agent.<domain> may query _mcp.<domain> for the depth [AGENT-AID] deliberately leaves out.¶

The keys bind different things, which is why publishing both is not a duplication. The k of [AGENT-AID] binds a key to an endpoint, and proves control of the server that answers. The _alter and _org-alter records bind a key to a ~handle or to a legal entity, with a log root and a revocation commitment. One proves the door is the right door. The other says who owns the building.¶

Revision 01 of this document made the first of these points, in a section titled "Relationship to Agent Identity and Discovery (AID)". Revision 02 dropped the prose and kept the reference, and revisions 03 and 04 dropped both. That was an error and this revision reverses it.¶

[DNS-AID] specifies discovery of AI agents through DNS. It carries connectivity in SVCB records with TXT as a fallback, an alpn parameter that may hold the transport, an application-layer agent protocol, or both (alpn=mcp,h2,h3), and an optional bap parameter for the agent protocol alone, so that consumers can match on the agent protocol without parsing transports out of alpn. It predates this document's treatment of the same question and is the most substantial neighbouring work in the space. It operates over unicast DNS, as this document does, and the two are complementary rather than competing. DNS-AID addresses agent discovery generally, while this document addresses one protocol family in depth and adds the identity records that carry a signed envelope.¶

[MDNS-AGENT] specifies zero-configuration discovery over mDNS and DNS-SD. Its proto parameter names the agent protocol family, and it defines exactly one value for it, proto=a2a, saying that A2A is the only instantiation included in that document and leaving other agent protocols to future work. It defines no registry for the key. The slot this document would occupy in that vocabulary, proto=mcp, is therefore open rather than taken, and the two documents agree on what the key MEANS, which is the part that matters for an implementer reading both. Its Section 7 puts discovery across links out of scope while noting that existing wide-area and aggregated DNS-SD facilities can provide it without changing the descriptor, so its scope is not permanently local. No record is ever shared with this document's records regardless, because the owner names and service types differ.¶

[MDNS-ARCHITECT] publishes registered agents through standard A, SVCB, and SRV records at a registry, with a worked SVCB example that puts the agent protocol and the transport side by side (SVCB 0 mcp ... alpn=h2). It is a third reading of the same axis, and it strengthens the case for a shared distinction rather than weakening it.¶

[SAIP] is a neighbour to the identity records of this document rather than to its service-discovery record. It publishes an agent's Ed25519 public key in an underscore-prefixed TXT record at _saip.<domain>, with a leading v= version tag, deriving its pattern, as this document does, from DKIM [RFC6376] and SPF [RFC7208]. Its Section 10, "Trust and Attestation Discovery", specifies that publication and the discovery of the key from it, so the two documents are solving the same problem at the same layer: how a verifier finds the key that a signature must check against. The bindings differ. The _saip record binds a key to a vendor's declared network footprint, carrying authorised ASNs, authorised IP prefixes, and an expiry. The _alter record of Section 7 binds a key to a ~handle, with a log root and a revocation commitment. A deployment that wanted both could publish both, and nothing in either record's owner name or field set collides with the other.¶

These drafts would serve their readers better with one vocabulary. An implementer reading two of them should not have to hold two meanings for one key in their head.¶

The neighbouring drafts keep the agent protocol and the transport apart, each in its own way. [MDNS-AGENT] defines proto as the agent protocol identifier and never as a transport, and carries the connection data in the SRV record instead. [MDNS-ARCHITECT] defines no proto key at all, and keeps the agent protocol distinct from the wire protocol inside its SVCB record, the one in the mcp position and the other in alpn. Neither of them puts a transport where the agent protocol belongs. Revision 01 of this document did exactly that, so revision 01 is the odd one out, and it is the reading that has to give way.¶

The alignment adopted here is therefore one parameter for the agent protocol family and a second, separate parameter for the transport or binding. The transport parameter is named transport because the Model Context Protocol specification already calls this axis by that name, so an MCP implementer meets one term, not two.¶

This document keeps a third axis separate again. transport names the binding (streamable-http, sse), and the wire protocol underneath it (h2, h3) is left to the alpn service parameter of SVCB and HTTPS records [RFC9460], where DNS already carries it. [DNS-AID] permits alpn to hold the transport, the agent protocol, or both, and a reader of that draft and this one should know that this document does not overload alpn in the same way. Three axes are made explicit here rather than two, and the third is not carried in TXT at all.¶

4. Why TXT rather than SVCB

The neighbouring drafts carry their connectivity data in SVCB records, with TXT as a fallback. This document is the reverse, TXT is the primary carrier. The reason is the payload, not a rejection of SVCB.¶

An _mcp record's value is almost entirely opaque tokens the resolver does not interpret, an endpoint URL, a public key, a capability list, an attestation scope. SVCB earns its place when a resolver or a connection library acts on the parameters directly, selecting a transport, following an alpn, choosing a port. It adds a registered SvcParamKey per field and a priority-ordered target list. For a record whose fields are consumed by an MCP client after resolution, and not by the resolver, that machinery is overhead without a matching benefit, and a TXT string is the lighter carrier.¶

Where a field IS resolver-actionable, this document defers to SVCB rather than duplicating it. The wire protocol lives in the SVCB or HTTPS alpn parameter [RFC9460], never in this TXT record, for exactly that reason. An operator who wants SVCB-based transport selection publishes an HTTPS record for the endpoint host in the ordinary way, and the two records coexist.¶

5. Record Format: _mcp.<domain> (Service Discovery)

This section defines the Discovery Record introduced in v01. It is given here in full. Revisions 03 and 04 incorporated it by reference, by a section number that named the wrong section, and a reader could not follow the pointer.¶

_mcp.<policy-domain>. IN TXT "<record-value>"

mcp-record      = version *( ";" SP field )
version         = "v=mcp1"
field           = url-field / proto-field / transport-field /
                  pk-field / epoch-field / cap-field /
                  attest-field / scope-field / priority-field /
                  ttl-field / ext-field / unknown-field

url-field       = "url=" https-uri
proto-field     = "proto=" proto-value
transport-field = "transport=" transport-value
pk-field        = "pk=" algo ":" base64url
epoch-field     = "epoch=" 1*DIGIT
cap-field       = "cap=" cap-csv
attest-field    = "attest=" attest-csv
scope-field     = "scope=" scope-csv
priority-field  = "priority=" 1*DIGIT
ttl-field       = "ttl=" 1*DIGIT
ext-field       = "ext=" https-uri
unknown-field   = token "=" *qtext

https-uri       = "https://" *uri-char
uri-char        = %x21-3A / %x3C-7E
                ; printable ASCII, excluding ";" (%x3B) and SP.
                ; A URI carries no raw space, so admitting one here
                ; would let a url= value swallow the field after it.
qtext           = %x20-3A / %x3C-7E
                ; printable ASCII and SP, excluding ";" (%x3B),
                ; which delimits one field from the next.  A value
                ; that could contain ";" could swallow the delimiter.
algo            = "ed25519"
base64url       = 1*( ALPHA / DIGIT / "-" / "_" )
proto-value     = token
transport-value = token
cap-csv         = cap-token *( "," cap-token )
cap-token       = token
attest-csv      = attest-token *( "," attest-token )
attest-token    = token
scope-csv       = scope-token *( "," scope-token )
scope-token     = "tools" / "resources" / "prompts" /
                  "sampling" / "identity" / token
token           = 1*( ALPHA / DIGIT / "-" / "_" )

_mcp.example.com. IN TXT (
  "v=mcp1; url=https://mcp.example.com; "
  "pk=ed25519:LongBase64UrlEncodedPublicKeyValueHere; "
  "epoch=3; attest=employ,contract,alumnus; scope=tools,identity"
)

6. Record Format: _org-alter.<domain> (Identity Bootstrap)

This section defines the Org-Identity Record introduced in v02. It is given here in full, for the same reason Section 5 is.¶

_org-alter.<policy-domain>. IN TXT "<record-value>"

orgalter-record = version *( ";" SP field )
version       = "v=alter1"
field         = org-field / entity-field / entity-type-field /
                founded-field / regions-field / regulated-field /
                bootstrap-field / mcp-policy-field /
                epoch-field / pk-field / attest-field /
                ext-field / unknown-field

org-field        = "org=" text-value
entity-field     = "entity=" registry ":" text-value
entity-type-field = "entity-type=" text-value
founded-field    = "founded=" date-fullyear
regions-field    = "regions=" region-csv
regulated-field  = "regulated=" framework-csv
bootstrap-field  = "bootstrap=" https-uri
mcp-policy-field = "mcp-policy=" policy-token
epoch-field      = "epoch=" 1*DIGIT
pk-field         = "pk=" algo ":" base64url
attest-field     = "attest=" attest-csv
ext-field        = "ext=" https-uri
unknown-field    = token "=" *qtext

registry      = "abn" / "acn" / "ein" / "ch" / "cro" / "lei" / token
text-value    = 1*qtext
qtext         = %x20-3A / %x3C-7E
              ; printable ASCII and SP, excluding ";" (%x3B),
              ; which delimits fields.  A legal entity name carries
              ; spaces, so VCHAR alone cannot express one.
region-csv    = region-token *( "," region-token )
region-token  = 2ALPHA   ; ISO 3166-1 alpha-2
framework-csv = framework-token *( "," framework-token )
framework-token = "disp" / "itar" / "ear" / "hipaa" / "gdpr" /
                  "soc2" / "iso27001" / "iso42001" / "essential8" /
                  "aprs" / token
policy-token  = "open" / "refuse-automated" / "refuse-tenant" /
                "refuse-all" / token
date-fullyear = 4DIGIT [ "-" 2DIGIT [ "-" 2DIGIT ] ]
algo          = "ed25519"
base64url     = 1*( ALPHA / DIGIT / "-" / "_" )
https-uri     = "https://" *uri-char
uri-char      = %x21-3A / %x3C-7E
              ; printable ASCII, excluding ";" (%x3B) and SP.
              ; A URI carries no raw space.
token         = 1*( ALPHA / DIGIT / "-" / "_" )

org=Alter Meridian Pty Ltd
org=Red Group Pty Ltd
org=International Business Machines Corporation

entity=abn:54696662049

regions=AU,NZ,SG

7. Record Format: _alter.<domain> (Envelope Publication)

This section defines the new Envelope Publication record introduced in v03. The record publishes the seven required fields of the ALTER identity envelope (binding a ~handle to its public key, IdentityLog root, inception timestamp, revocation commitment, and detached signature, under a protocol version) at an underscore- prefixed label under the handle's hosting zone. The envelope schema admits an optional caveats array which is never carried in DNS; this document specifies no transport and no vocabulary for it, so under this document the array is always empty and the signature is computed over an empty array (Section 7.4). The signature algorithm is not a separate field. It is derived from the pk= prefix via the registry of Section 16.7, and enters the signing input under that derivation.¶

_alter.<zone>. IN TXT "<record-value>"

; Publisher emission.  A conformant publisher MUST emit this form.
alter-record   = version ";" SP handle-field
                 ";" SP pubkey-field
                 ";" SP ilr-field
                 ";" SP ts-field
                 ";" SP rev-field
                 ";" SP sig-field
                 *( ";" SP unknown-field )

; Resolver acceptance.  A conformant resolver MUST accept this form.
; The version field stays first.  The remaining six required fields
; may arrive in any order, each exactly once, interleaved with any
; number of unknown fields.
alter-accepted = version *( ";" SP accepted-field )
accepted-field = handle-field / pubkey-field / ilr-field
               / ts-field / rev-field / sig-field
               / unknown-field

version        = "v=alter1"
handle-field   = "h=" handle
pubkey-field   = "pk=" algo ":" base64url
ilr-field      = "ilr=" base64url     ; SHA-256 of IdentityLog root
ts-field       = "ts=" 1*DIGIT        ; inception_ts, Unix seconds
rev-field      = "rev=" base64url   ; SHA-256 of revocation pre-image
sig-field      = "sig=" base64url     ; Ed25519 detached signature
unknown-field  = token "=" *qtext

handle         = "~" 1*( ALPHA / DIGIT / "-" / "_" )
                 [ "." "bot" ]
               / "~cc-" 1*( ALPHA / DIGIT / "-" / "." )
algo           = "ed25519"
base64url      = 1*( ALPHA / DIGIT / "-" / "_" )
token          = 1*( ALPHA / DIGIT / "-" / "_" )
qtext          = %x20-3A / %x3C-7E
               ; printable ASCII and SP, excluding ";" (%x3B),
               ; which delimits one field from the next.

pk=ed25519:<base64url-no-pad-32-bytes>

8. DNSSEC Requirement

The zone publishing an _alter.<domain> envelope record MUST be DNSSEC-signed per [RFC4033], [RFC4034], and [RFC4035]. Authoritative servers MUST respond with valid RRSIG coverage for the TXT RRset. Recursive resolvers handling queries for the envelope RRset MUST perform DNSSEC validation and MUST set the AD (Authenticated Data) bit on the response delivered to the stub client.¶

Stub clients (MCP clients, alter-runtime daemons, onboarding wizards, recognition verifiers) consuming _alter.<domain> envelope records MUST reject any response that lacks a set AD bit or that fails local RRSIG verification when operating in validating-stub mode. An envelope obtained over an unvalidated DNS path is not an envelope; it is unauthenticated TXT content. Treating it otherwise is a downgrade vulnerability (Section 14.1).¶

This requirement is specific to _alter.<domain> records. DNSSEC is RECOMMENDED but not REQUIRED for _mcp.<domain> and _org-alter.<domain> records in this revision, for backward compatibility with v01 and v02 deployments. Future revisions of this document MAY promote DNSSEC to REQUIRED for the other two records once deployment data justifies the promotion.¶

9. DANE TLSA Pin

The MCP endpoint associated with a published envelope MUST carry a DANE TLSA resource record [RFC6698] binding the endpoint's leaf TLS certificate or SubjectPublicKeyInfo to the zone. The TLSA record MUST be published at:¶

_443._tcp.mcp.<zone>. IN TLSA <usage> <selector> <matching-type>
                              <cert-association-data>

Recommended parameters:¶

• Usage field. 3 (DANE-EE), pin the end-entity certificate directly, with no CA chain reliance. A publisher that explicitly requires CA-chain validation MAY use 1 (PKIX-EE) instead. Publishers MUST NOT use 0 (PKIX-TA) or 2 (DANE-TA) for the envelope organ; the trust basis of the envelope is the end-entity leaf.¶

• Selector field. 1 (SPKI), pin the SubjectPublicKeyInfo so that certificate rotations preserving the keypair do not invalidate the record. Selector 0 (full certificate) MAY be used but requires more frequent TLSA republication.¶

• Matching-type field. 1 (SHA-256). Matching type 2 (SHA-512) is reserved for future revisions.¶

Clients establishing an MCP session at https://mcp.<zone>/ in conjunction with a resolved envelope MUST fetch and validate the TLSA record, MUST abort the TLS handshake on mismatch, and MUST NOT fall back to PKIX-only validation on TLSA failure.¶

The TLSA requirement is scoped to envelopes whose MCP session establishment is triggered by the resolved envelope (i.e. when the envelope resolution and the subsequent MCP session are part of a single recognition transaction). MCP clients that do not resolve an envelope (e.g. v01-only clients consuming only _mcp.<domain>) are out of scope for this requirement and continue to operate under v01 rules.¶

10. IdentityLog Cross-Reference

The ilr= field of the _alter.<domain> record carries a base64url-no-pad SHA-256 digest of an IdentityLog root claimed by the publisher to have been witnessed at envelope creation.¶

Revision 04 of this document required resolvers to cross-reference that value against one of four named witness surfaces, and treated a failure to do so as fatal to recognition. That requirement is WITHDRAWN. It named surfaces that do not exist, and it would not have achieved what it claimed even if they did.¶

What ilr= can and cannot establish, stated plainly:¶

• A resolver that has an independent view of the log MAY confirm that the root named in ilr= is one the log actually published.¶

• A resolver MUST NOT conclude from ilr= that the envelope in front of it is recorded in that log. The field carries a root and nothing else. It carries no leaf hash, no audit path, and no proof of inclusion, so no resolver can perform that check, and Section 14.3 sets out the forgery this permits.¶

• A resolver MUST NOT treat the presence, absence, or successful lookup of ilr= as a defence against envelope substitution.¶

The field is retained so that withdrawing its security claim does not also force a breaking change to the wire format. It is advisory.¶

The log's leaf hashing, tree construction, cadence, and witness protocol are not specified by this document, and a reader should not assume a specification exists.¶

The revocation check also crosses the IdentityLog. Where a resolver has access to the log's revocation surface, it MUST treat the envelope as revoked, and MUST NOT honour it regardless of the freshness of the TXT RRset, if a pre-image whose SHA-256 equals the rev= field has been revealed there.¶

A reader implementing from this document alone has no such access. The surface that carries reveals is not specified here and is not published anywhere a reader can fetch, so revocation status cannot presently be established from DNS alone. This is a gap in the mechanism, not a property of it, and it is stated rather than papered over.¶

11. alter: URI Scheme Cross-Reference

An IANA-registered URI scheme alter: provides a dispatchable surface for ~handle references: operating-system URI handlers (xdg-mime, LSHandlers, Windows registry, Android intent-filter) invoke a resolver that retrieves the envelope from DNS and verifies it per Section 12.3, falling back to the HTTPS .well-known surface where the DNS record is absent.¶

Registration is provisional per [RFC7595] Section 3. The full registration body, scheme syntax, semantics, encoding considerations, interoperability and security considerations, author and change controller, is submitted to IANA separately from this document. The IANA request is in progress as of the publication date of this revision; the final registration reference will be substituted when available.¶

Handlers invoked via alter: URIs MUST perform full envelope verification, DNSSEC validation (Section 8), and DANE TLSA binding (Section 9) when establishing any HTTPS session, before acting on any content or directive derived from the envelope.¶

12. Discovery and Bootstrap Procedures

The three procedures below are normative and complete. None of them is a summary of an algorithm specified elsewhere.¶

Revisions 03 and 04 of this document did not restate the first two. They pointed at them by section number, in a document the reader was expected to fetch separately, and every one of those pointers named the wrong section. Nothing is incorporated by reference now. The formats these procedures consume are given in full in Section 5 and Section 6, and the procedures themselves are below.¶

13. Caching

The rules for the first two records are those of [DNSDISC-01] and [DNSDISC-02], unchanged, restated here so that a reader need not fetch either document to cache correctly.¶

For _mcp.<domain>, clients SHOULD cache the parsed record metadata for the duration of the DNS TTL. Where the record carries a ttl field, clients MAY extend the metadata cache to that duration, but MUST re-validate the underlying DNS record when the DNS TTL expires. A client that has connected to an MCP server and verified its pk SHOULD cache the verified key binding and re-validate it on subsequent connections, which is trust on first use with periodic re-verification against DNS.¶

For _org-alter.<domain>, records SHOULD be cached for the duration of the DNS TTL. An onboarding wizard typically reads the record once at install time and persists the resolved state locally, so the DNS record need not be re-read on every invocation. Wizards MAY re-read it on operator request, on epoch change detected by a periodic background poll, or on identity verification failure.¶

_alter.<domain> records SHOULD be cached for the duration of the DNS TTL. Resolvers MUST NOT serve stale envelope TXT past the RRset TTL unless they are themselves validating caches and can re-confirm RRSIG coverage on each serve. Recognition verifiers MAY cache successful verification results locally for a short interval (bounded above by the RRset TTL or 3600 seconds, whichever is smaller) to amortise the cost of repeated JCS and Ed25519 operations. A resolver that is able to perform the revocation check at all (Section 10 sets out why one working from this document alone is not) MUST re-run it on each recognition event, not on each cache refresh.¶

14. Security Considerations

(Security considerations from v01 and v02 are retained. Additional considerations introduced by the envelope layer are below.)¶

15. Privacy Considerations

(Privacy considerations from v01 and v02 are retained. Additional considerations introduced by the envelope layer are below.)¶

16. IANA Considerations

17. Examples

This section provides non-normative examples of Envelope Records for common deployment scenarios.¶

_alter.example.com. 3600 IN TXT (
  "v=alter1; h=~alice; "
  "pk=ed25519:<EXAMPLE-pubkey-32B-base64url>; "
  "ilr=<EXAMPLE-identitylog-root-32B-base64url>; "
  "ts=1729123456; "
  "rev=<EXAMPLE-revocation-hash-32B-base64url>; "
  "sig=<EXAMPLE-ed25519-signature-64B-base64url>"
)

_alter.example.org. 3600 IN TXT "v=alter1; h=~alice; pk=..."
_alter.example.org. 3600 IN TXT "v=alter1; h=~bob; pk=..."
_alter.example.org. 3600 IN TXT "v=alter1; h=~carol.bot; pk=..."

_mcp.example.com.      IN TXT "v=mcp1; url=https://mcp.example.com/"
_org-alter.example.com. IN TXT "v=alter1; org=Example Org; ..."
_alter.example.com.     IN TXT "v=alter1; h=~alice; "
                               "pk=ed25519:...; ilr=...; "
                               "ts=...; rev=...; sig=..."
_443._tcp.mcp.example.com. IN TLSA 3 1 1 <sha256-of-spki>

_alter.example.com. 3600 IN TXT (
  "v=alter1; h=~cc-example-model; "
  "pk=ed25519:...; ilr=...; ts=...; rev=...; sig=..."
)

18. Interoperability with Earlier Record Generations

A domain that publishes only a v01 _mcp.<domain> record continues to work with all v01, v02, and v03 clients. Where that record carries a transport value in proto, a client conforming to this revision reads it under rule 3 of Section 5.3.5 and reaches the same endpoint over the same transport. No republication is required for the record to keep working, and none is required for a record carrying a proto value that no revision recognised, because such a record was skipped by v01 clients and is skipped by these.¶

A domain that publishes _mcp.<domain> and _org-alter.<domain> (v02) continues to work with v02 clients and with clients conforming to this document, unchanged. A client conforming to this document may additionally query _alter.<domain> and MUST handle its absence gracefully, which is the common case and the recommended one: publication of that record is NOT RECOMMENDED (Section 7.1.1), so a conforming client should expect to find it absent and MUST NOT treat its absence as an error.¶

A domain that publishes all three records benefits from:¶

• Service discovery via _mcp.<domain> (v01).¶

• Organisational identity bootstrap via _org-alter.<domain> (v02).¶

• Individual identity recognition via _alter.<domain> (v03).¶

• DNSSEC-authenticated envelope delivery (Section 8).¶

• DANE TLSA binding on the MCP endpoint (Section 9).¶

A domain that publishes only _alter.<domain> (envelope-only, no MCP server, no organisational record) is permitted by the grammar. Revision 04 called it the appropriate configuration for a Sovereign-tier individual. That recommendation is WITHDRAWN, and the configuration is NOT RECOMMENDED per Section 7.1.1.¶

The three records are orthogonal along their semantic axes but share the zone's DNSSEC trust root. A resolver conforming to this document that resolves any subset of the three records treats each resolution as independent and does not fail the resolution of one record because another is absent or malformed.¶

19. Implementation Status

This section records the status of known implementations at the time of publication, per [RFC7942].¶

This section was materially wrong in revision 04, which described a deployment that does not exist. It is rewritten here against the zone as it actually resolves and the code as it is actually written. What follows is deployed, and nothing else is claimed.¶

Deployed and resolvable in the truealter.com zone:¶

• _mcp.truealter.com, carrying v=mcp1 with url, proto, pk, epoch, and cap. This is the only record in the zone that a resolver can verify against this document today.¶

• _alter.truealter.com, carrying v=alter1. It does not carry the envelope field set (h, ilr, ts, rev, sig) and is therefore NOT an instance of the Envelope Record of Section 7. No conformant Envelope Record is published in this zone.¶

Implemented in code, and NOT exercised against any conformant envelope, because none is published:¶

• A resolver and verification library implementing the JCS signing-input construction of Section 7.4 and the Ed25519 signature check. It conforms to this revision on the signing input and the signature check. It does NOT conform on two steps of the recognition procedure of Section 12.3: it still treats the IdentityLog cross-reference as fatal, where step 9 now says a resolver MUST NOT abort, and it still fetches caveats over HTTPS, which step 11 now places out of scope. Those two steps follow revision 04, and this document does not claim otherwise.¶

NOT deployed, and stated plainly because revision 04 claimed otherwise:¶

• _org-alter.truealter.com does not resolve. NXDOMAIN.¶

• _443._tcp.mcp.truealter.com does not resolve, so no DANE TLSA pin is published and the DANE binding of Section 9 is untested in deployment.¶

• truealter.com is NOT DNSSEC-signed. It publishes no DNSKEY and the parent zone holds no DS. Section 8 of this document is a MUST, and the operator does not yet meet it. No Envelope Record can be relied upon in this zone until it does.¶

• There is no signed-tree-head federation and no witness-mirror network. Revision 04 described four independent witness surfaces including an on-chain anchor contract. None of them exist.¶

• No inclusion proof is generated or checked anywhere. The ilr field of Section 7.2 is published as a bare root with no proof attached. Section 14.3 sets out what that permits.¶

The Envelope Record of Section 7 therefore has NO conformant deployment at the time of writing, in this zone or any other known to the author. It is specified, implemented in a verifier, and unpublished.¶

The deployed _alter.truealter.com record carries none of the envelope fields that Section 7 defines, so it is not an instance of the Envelope Record and a resolver MUST NOT treat it as one.¶

20. References

Appendix A. Recognition Pseudocode

The following pseudocode illustrates the combined recognition procedure defined in Section 12.3. It is non-normative; the normative procedure is the twelve-step algorithm in the body of this document.¶

function recognise_envelope(handle, zone):
    # Step 1-2: Query + DNSSEC
    response = dns_query("_alter." + zone, type=TXT, prefer=DoH)
    if not response.ad_bit and not local_rrsig_validate(response):
        raise UnauthenticatedResponse

    # Step 3-5: Chunk reassembly + handle disambiguation + fields
    records = [parse_alter_record(rr) for rr in response.rrset]
    record = find(records, lambda r: r.h == handle)
    if record is None or record.v != "alter1":
        raise RecordNotFound
    for f in ["h", "pk", "ilr", "ts", "rev", "sig"]:
        if not hasattr(record, f):
            raise MalformedRecord

    # Step 6-7: Envelope reconstruction + JCS.  Keys are the TXT
    # field names; every value stays a string, ts included.
    alg = signature_alg_for_prefix(record.pk)   # Signature Algorithm
                                               # Registry
    envelope = {
        "v": record.v,
        "h": record.h,
        "pk": record.pk,
        "ilr": record.ilr,
        "ts": record.ts,
        "rev": record.rev,
        "caveats": [],
        "signature_alg": alg,
    }
    signing_input = jcs_canonicalise(envelope)

    # Step 8: signature verify under the derived algorithm
    if not signature_verify(alg, record.pk, record.sig,
                            signing_input):
        raise SignatureInvalid

    # Step 9: IdentityLog cross-ref. ADVISORY. Confirms the ROOT
    # was published; cannot confirm THIS envelope is under that
    # root, so a failure annotates and never aborts. See 12.3.
    ilr_root_seen = identitylog_root_published(record.ilr)

    # Step 10: DANE TLSA (if establishing MCP session)
    if establishing_mcp_session(zone):
        tlsa = dns_query("_443._tcp.mcp." + zone, type=TLSA)
        if not tlsa_matches_endpoint(tlsa, "mcp." + zone):
            raise TLSAFailure

    # Step 11: Caveats are OUT OF SCOPE for this document (Sec 10.3).
    # No transport and no vocabulary are specified, so the verified
    # envelope carries none and the signature is checked over [].
    caveats = []

    # Step 12: Revocation
    if identitylog_revocation_revealed(record.rev):
        raise EnvelopeRevoked

    return VerifiedEnvelope(record, caveats, ilr_root_seen)

Appendix B. Document History

draft-morrison-mcp-dns-discovery-05 (July 2026):¶

Corrections to defects in -04, and the completion of a document that had been claiming for two revisions to stand on its own while deferring half of itself to documents it cited by the wrong section number. No new mechanism is introduced in this revision.¶

• Restores the citation of [AGENT-AID], which revision 01 carried and revisions 02 through 04 lost. It is the closest neighbour to the _mcp record and it published its shape first, and a reader of both documents should be told so by this one. Section 3 also distinguishes it from [DNS-AID], a different draft with the same acronym, and states how the two compose: [AGENT-AID] is intentionally small and defers to protocol-specific mechanisms after discovery, and this document is what its mcp protocol token defers to.¶

• Makes the stand-alone claim true. All three record formats are now given in full (Section 5, Section 6, Section 7), as are all three procedures (Section 12.1, Section 12.2, Section 12.3) and the caching rules (Section 13). Revisions 03 and 04 incorporated the first two records and their procedures by reference to revisions 01 and 02. Nothing is incorporated by reference now, and no reader of this document needs to fetch another one in the series.¶

• Separates the agent protocol family from the transport binding in the _mcp record. Revision 01 used proto for the transport. proto now names the protocol family, transport names the binding, and transport is the term because the Model Context Protocol specification already calls this axis by that name. The alignment is with the neighbouring DNS agent-discovery drafts (Section 3), so an implementer reading two of them does not have to hold two meanings for one key.¶

• States how to read an _mcp record published under revision 01 (Section 5.3.5), including the case revision 01 left unusable. A proto value that names neither the protocol family nor a known transport causes the record to be SKIPPED, exactly as revision 01 required, and is NOT read as a protocol family with the default transport applied. Widening the rule would newly accept records no client could ever use.¶

• Reconciles the _mcp grammar with its prose. The grammar admits any token in proto and transport, because a resolver must parse a revision 01 record before it can decide what to do with it. Syntactic admissibility is not definition. proto has exactly one defined value and transport exactly three, the two sets are disjoint, and so the claim that no value is defined for both fields is now true, where in earlier drafting it was not.¶

• Corrects a grammar defect inherited from revision 02. The _org-alter record defined org, entity and entity-type over 1*VCHAR, which excludes the space, so the grammar could not derive the values revision 02's own examples published. A text-value rule admitting SP, and excluding only the field separator, replaces it (Section 6.2). This is the first revision to carry that format as its own normative text, so it is the first that had to be able to derive its own examples.¶

• Stops every value rule in all three grammars from running past the end of its own field. unknown-field and https-uri were defined over *VCHAR, and VCHAR includes the semicolon, so a single field could derive the whole rest of the record and consume the fields after it. Two character classes replace it, because the two kinds of value have different needs: qtext admits SP and excludes the semicolon, for a human-readable name; uri-char excludes both, for a URI, which carries no raw space and would otherwise run past its own end. A class that is right for a name is not thereby right for a URI.¶

• States the exclusion that ABNF cannot express (Section 5.2). unknown-field matches only a field name this document does not define, so a defined field whose value is malformed is a malformed record and MUST be discarded, never re-read as an unknown field. Without the rule the alternation swallows every defect it exists to exclude, because token matches a defined name as readily as an undefined one. The split-on-semicolon parse in Section 12.1 already behaved this way, so no implementation changes; the grammar now says what the parser was always doing.¶

• Requires a publisher emitting transport to emit proto=mcp with it (Section 5.3.5). A revision 01 client ignores the transport field it does not know and reads proto, so transport=streamable-http; proto=sse was one record that resolved to two different transports depending on which revision read it.¶

• Adds Section 3 and Section 4, which situate this document against [DNS-AID], [MDNS-AGENT], [MDNS-ARCHITECT], and [SAIP], and say why the payload here is carried in TXT rather than SVCB.¶

• Corrects the IANA registry table (Section 16.1), which cited the three labels to sections of revisions 01, 02 and 03, by numbers that were wrong in any case. Each label is now cited to the section of this document that defines it.¶

• Rewrites Section 19 (Implementation Status). The -04 section was false on four counts. It claimed an _org-alter record that does not resolve, a DANE TLSA pin that does not resolve, an envelope record exercising the Section 7 field set where the deployed record carries none of those fields, and a four-surface signed-tree-head witness federation, including an on-chain anchor contract, none of which exists. The section now states only what is deployed, and states plainly what is not.¶

• Corrects Section 14.3 defence (1). The -04 text claimed that ilr= defeats envelope substitution. It does not. ilr= is a bare root with no inclusion proof, so a forged envelope naming any genuinely witnessed root passes every specified check. The false claim is withdrawn and the limit is stated.¶

• Fixes the signing input of Section 7.4, which no third party could have interoperated with. The keys are now the TXT field names, v is inside the signed object, and ts is a JSON string of the wire digits rather than a JSON number. The construction is now normative and exhaustive.¶

• Derives signature_alg from the pk= algorithm prefix via the registry of Section 16.7, rather than injecting Ed25519 as an implicit constant into the signed bytes. Taken alone this change is byte-compatible, because for pk=ed25519: the derivation yields the string the constant supplied. The revision as a whole is NOT byte-compatible: the signing-input repair above changes the signed bytes deliberately, so an envelope signed under -04 does not verify under -05. No envelope signed under -04 exists.¶

• Relaxes the multi-string rule of Section 7.6, which prohibited splitting within a key-value pair and thereby made any field longer than 255 octets unrepresentable in a TXT record. Resolvers already concatenate before parsing, so the prohibition bought nothing and cost the ability to carry a large signature at all.¶

• Marks publication of a per-individual envelope in DNS as NOT RECOMMENDED (Section 7.1.1). A zone MUST NOT publish envelopes for more than one handle at one owner name, because a single query there returns the whole set, which is a membership roll that no one listed in it can consent to on behalf of the others. The envelope format, its signing input and its verification procedure are not implicated.¶

• Withdraws the IdentityLog witness federation, everywhere it was asserted. Revision 04 defined it in the terminology, required it normatively in Section 7.3.4 and Section 10, made it fatal at step 9 of the recognition procedure, and named four witness surfaces including an on-chain anchor contract. None of those surfaces exist. The ilr= field is retained, so the wire format does not break, but the cross-reference is now advisory and the document states plainly what the field can and cannot establish.¶

• States, rather than papers over, the fact that a reader implementing from this document alone cannot check revocation. The surface that carries reveals is not specified here and is not published anywhere fetchable.¶

• Repairs every internal cross-reference, and repairs the class rather than the instances. The -04 prose was written against an older section numbering and never renumbered, so a reader following a pointer to the Envelope Record landed on the DNSSEC section. Every reference to a section of this document is now a symbolic anchor that cannot go stale under renumbering. The only literal section numbers that remain in the prose point into other documents.¶

• Corrects all four references into [DNSDISC-01] and [DNSDISC-02]. Every one of them named the wrong section. The _mcp record format is Section 5 of [DNSDISC-01] and was cited as Section 3; its discovery procedure is Section 6 and was cited as Section 4. The _org-alter record format is Section 6 of [DNSDISC-02] and was cited as Section 4; its bootstrap procedure is Section 7 and was cited as Section 6. A reader who followed any of them arrived at the wrong section of the right document.¶

• Restates the discovery procedure (Section 12.1), the identity bootstrap procedure (Section 12.2), and the caching rules for both records (Section 13) in full. Revisions 03 and 04 claimed to stand alone while deferring three procedures to documents they cited incorrectly. The record formats are restated in full as well, so nothing at all is now incorporated by reference, and both earlier revisions carry a reference entry, which neither had.¶

• Gives the _alter record two ABNF grammars (Section 7.2), one for publisher emission and one for resolver acceptance. Revision 04 gave the publisher grammar alone, then required resolvers to accept orderings that grammar cannot derive. The field-cardinality rule, which ABNF cannot express, is stated normatively beside it, and the v-first rule is reconciled with the ordering freedom rather than contradicting it.¶

• Reconciles Section 14.6 with step 5 of Section 12.2. Revision 04 said the _mcp and _org-alter keys MAY differ, while the bootstrap procedure it incorporated by reference required them to MATCH and required a wizard to refuse on mismatch. The bootstrap rule stands. The envelope key of Section 7 is the key that is genuinely distinct, and it alone MAY differ from the other two.¶

• Removes two hand-written reference sections that duplicated the generated ones and cited five documents with no reference entry.¶

• Corrects the field count throughout. Seven fields are REQUIRED; the -04 prose said five in three places.¶

draft-morrison-mcp-dns-discovery-03 (April 2026):¶

Editorial corrections (retiring -02):¶

• Removes the third-party-domain worked example used in -02 and replaces all instances with [RFC2606] reserved example-domain forms; no third-party operational domain appears in any illustrative DNS record in this revision.¶

• Strips city and locality fields from the author front-matter block, retaining only name, organisation, and email per editorial policy.¶

Substantive additions:¶

• Adds the _alter.<domain> Envelope Record (Section 7).¶

• Defines v, h, pk, ilr, ts, rev, sig fields for the new record.¶

• Introduces a mandatory DNSSEC validation requirement for _alter.<domain> responses (Section 8).¶

• Introduces a mandatory DANE TLSA [RFC6698] pin on the MCP endpoint (Section 9) for envelope-triggered MCP sessions.¶

• Adds the IdentityLog cross-reference requirement (Section 10). Revision 05 withdraws it; see the -05 entry above.¶

• Adds a provisional alter: URI scheme cross-reference per [RFC7595] (Section 11).¶

• Adds the envelope recognition procedure (Section 12.3), a twelve-step algorithm.¶

• Adds IANA registration for _alter underscore-prefixed label (Section 16.1) and the independent v=alter1 envelope version namespace.¶

• Adds a Signature Algorithm Registry (Section 16.7) with initial value ed25519.¶

• Adds Security Considerations for DNSSEC downgrade, TLSA pin rotation, envelope substitution, revocation opacity, clock skew, cross-record key consistency, and passive-stream coupling.¶

• Adds Privacy Considerations for public handle disclosure, DNS query metadata, and revocation unlinkability.¶

• Adds Examples for minimal envelope, multi-handle zone, full ALTER zone with all three records, and Instrument-tier handle.¶

• Adds Implementation Status entry for the envelope reference implementation.¶

• Incorporated the v01 _mcp.<domain> and v02 _org-alter.<domain> record specifications by reference, leaving them unchanged. Revision 05 restates both in full; see the -05 entry above.¶

draft-morrison-mcp-dns-discovery-02 (April 2026):¶

• Adds the _org-alter.<domain> Org-Identity Record.¶

• Defines org, entity, entity-type, founded, regions, regulated, bootstrap, mcp-policy, epoch, pk, attest, ext fields for the organisational record.¶

• Adds the Identity Bootstrap procedure.¶

• Adds IANA registration for _org-alter underscore-prefixed label.¶

• Adds version tag v=alter1 (org-alter namespace) and registry namespace and framework token registries.¶

• Adds Examples for minimal, full, regulated (DISP), and multi-regulator deployments.¶

• Adds Implementation Status entry for the orgalter_discover reference library.¶

• v01 _mcp.<domain> record specification is incorporated by reference and remains unchanged.¶

draft-morrison-mcp-dns-discovery-01 (April 2026):¶

• Adds Identity Field Theory grounding for epoch and scope.¶

• Refines security considerations for identity assurance decay.¶

• Refines privacy considerations for scope as a privacy boundary.¶

• Adds Coexistence section with SEP-1959, AID, A2A.¶

• Adds Implementation Status section.¶

draft-morrison-mcp-dns-discovery-00 (April 2026):¶

• Initial submission.¶

• Defines _mcp.<domain> TXT record format with ABNF grammar.¶

• Defines discovery procedure with HTTPS fallback.¶

• Defines pk, epoch, attest, scope, cap, priority, ttl, and ext fields.¶

• Registers _mcp in the underscored DNS node name registry.¶

Contributors

Author's Address