The Scope Creep Automated reviews of new AI-related working group submissions

Review · November 22, 2026

Discovery of LLM API Services in Local Networks

draft-cui-dawn-local-llm-api-discovery -00View on Datatracker →

The Problem

A developer moves their laptop from the office to home. Both networks have a local LLM server. The AI coding tool still points at the office one, and someone has to edit the base URL by hand.

This draft defines a DNS-SD service type, _llm-api._tcp, so an LLM API can announce itself on the local link over mDNS. A client finds it, builds the base URL, and offers it to the AI application after the user approves.

How It Fits In

This is a DNS-SD profile. It adds no new protocol and extends no existing RFC. It depends on mDNS (RFC 6762), DNS-SD (RFC 6763), and TLS service identity (RFC 9525). It has no OAuth dependency.

The local network is one of the three DAWN charter use cases, and the charter names mDNS as a mechanism to build on. That part fits.

Two things don't fit as cleanly.

The charter says work items "will provide integrity of all the discovery information exchanged." This draft says it defines no integrity protection and leaves it to a future revision.

The charter is also about AI agents first. An LLM inference endpoint is an AI resource, but it is a different thing from an agent. The working group should decide whether it wants per-resource-type service names.

So how does this relate to the other mDNS draft? draft-jakab-dawn-agent-discovery-mdns already defines _agents._tcp with a proto key, subtypes per protocol, and keys shared with unicast DNS through draft-blwm-dawn-common-discovery-parameters. It lists local model servers among its use cases. This draft cites neither. It defines a second service type and a second set of TXT keys for the same job.

How It Works

There are four players: the LLM API service, an advertiser that publishes its DNS-SD records, a discovery client on the user's machine, and the AI application.

The advertiser publishes the usual PTR, SRV, TXT, and A/AAAA records. The TXT record carries six keys:

  • txtvers: the TXT format version
  • api: the API Profile, which says what kind of API is at the endpoint
  • protovers: the version of that API, when the profile requires one
  • scheme: http or https
  • base: the path prefix, such as /v1
  • auth: a hint, required or none

The client combines the scheme, SRV host, SRV port, and base path into a URL like https://llm-node.local:8443/v1. It shows that candidate to the user or a local policy. After approval it hands the URL and profile to the application.

One API Profile is defined: openresponses, pointing at the Open Responses specification. Chat Completions-style APIs are mentioned as future bindings.

Credentials, model names, and trust are all explicitly out of scope. Inference requests go straight from the application to the service.

Interoperability

The wire format is small and well specified. Two implementations would agree on the records and build the same URL. The problems are in what happens next.

Significant interoperability concerns

The user still has to configure the service by hand. Discovery delivers a base URL. The model identifier and the credential are "obtained separately." The one defined binding requires auth=required, so every discovered service needs a key the user has to go find and paste in. The draft removes one of three manual steps.

HTTPS with a .local name has no workable trust path. No public CA issues a certificate for llm-node.local. The server needs a private CA that the client already trusts, which is its own configuration step. The draft requires RFC 9525 verification and then says that verification is not enough to release credentials.

HTTP is permitted but unusable. Clients MUST NOT send credentials or sensitive data over plain HTTP. The only binding requires authentication, and prompts are sensitive data. Nothing can be sent to an http endpoint under these rules. Most home LLM servers run plain HTTP today.

Service identity verification is required and undefined. Step 7 says the service identity MUST be verified before credentials are sent. Approval MUST NOT rely solely on the advertised name. The draft gives the user nothing else to check. Any device on the Wi-Fi can advertise Office-LLM and collect API keys from whoever clicks approve.

auth=required doesn't say how. There is no way to tell a Bearer API key from OAuth. A client can't prompt for the right thing. An OAuth-protected service could point at Protected Resource Metadata and the hint would become unnecessary.

API Profiles have no registry. The IANA section registers the service name only. The draft says each binding needs an identifier, a version, and endpoint construction rules, and doesn't say where those get recorded. Two vendors will pick different strings for Chat Completions.

The only binding is not what local servers mostly speak. Ollama, LM Studio, vLLM, and llama.cpp are known for Chat Completions-style endpoints. That binding is deferred. The one that exists is a normative reference to a date-versioned spec outside any standards body.

protovers has no matching rule. Does a client that supports 2026-04-24 accept a later date? The draft doesn't say.

Mechanisms that could be removed

  • auth. It is a hint with no client behavior attached. For the only binding it is always required.
  • The approval and handoff steps. How a client shows candidates and writes application settings is local behavior. A paragraph in Security Considerations covers it.
  • "Does not make an inference request to verify discovery." A peer can't observe this.
  • The attachment-change rules. Respecting record lifetimes is already in RFC 6762. Not carrying a credential to a new service belongs in Security Considerations.
  • The whole service type. api=openresponses could be a proto value and a subtype under _agents._tcp, with base as the only new key.

Writing Quality

The writing is clear and concise. Thirteen pages is the right size. RFC 2119 language is used correctly, and the base path validation rules are precise.

The draft is honest about its gaps. It says plainly that integrity protection and identity binding are unspecified.

Some MUSTs are not testable. "The service identity MUST be verified" and "apply local destination restrictions" name an outcome with no procedure.

There are no informative references and no related work section. For a draft submitted to DAWN, not mentioning the existing DAWN mDNS draft is a real omission.

Background of the Author

Yong Cui (Tsinghua University) has 11 RFCs, published between 2009 and 2019, in softwires and DHCPv6. He is an experienced IETF author.

He has 21 active drafts and 55 expired ones. Most of the active ones are AI agent or LLM drafts across DAWN, NMRG, NMOP, and IDR. Three are in DAWN: this one, the MDI model, and cooperating discovery.

Questions Worth Raising

  1. Why a new service type? What does _llm-api._tcp do that a proto value and subtype under _agents._tcp can't?
  2. What does the user actually skip? If the API key and model name are still entered by hand, is typing the URL the problem worth a standards-track document?
  3. How does a client verify a .local service? Without an answer, the security requirements can't be met and implementations will ignore them.
  4. Where do API Profile identifiers get registered?
  5. Is an LLM inference endpoint in scope for DAWN? If yes, the common parameters draft should cover it.