The registry vacuum: who owns the name before the spec ships?

Sources cited in this piece · Official MCP Registry: registry.modelcontextprotocol.io · modelcontextprotocol/registry (quoted verbatim) · third-party directory a2a-registry.org (mechanics quoted from its own pages) · Forkast reporting 2026-10-10 (archived) · our own registry experience, dated 2026-09/10, first-hand

The A2A Agent Registry formal specification missed its October delivery target, and the vacuum it left is not sitting empty. As of this writing: at least five independent registry implementations ship with incompatible APIs; the official Model Context Protocol (MCP) Registry has frozen its v0.1 API and lists servers today; a third-party directory added seventy-five agents in the past seven days under a list-first, claim-later policy. The question “who owns the name” is being answered right now — not by a spec, but by whatever defaults ship first.

1. Two philosophies, already live

The official MCP Registry states its rule in one sentence: “The registry validates namespace ownership when publishing.” The namespace itself is rooted in something the publisher provably owns — io.github.domdomegg/my-cool-mcp requires logging in as domdomegg; me.adamjones/my-cool-mcp requires proving ownership of adamjones.me via a DNS or HTTP challenge. Validate first, then list. The listing is a claim that was checked at write time.

A third-party directory for A2A agents runs the opposite sequence: paste an agent URL — no account needed — and the entry goes live immediately; the purported owner is invited to come and claim it afterwards. The site’s own copy says it plainly: “Spot yours unclaimed? Sign in to claim it.” Verification exists (DNS TXT, GitHub association) but is decoupled from listing; unclaimed entries merely rank lower. List first, verify later. Until the owner shows up, the entry is a claim without a cost.

2. Why the vacuum matters: names harden

The formal A2A registry design — three layers (Agent Card, Publication Record, Trust-Authorization Overlay) and three operations — converged in the community more than a year ago, but was never ratified; in August the maintainers redirected discovery work toward two adjacent specifications, each with its own ratification still ahead. Meanwhile the implementations multiply: Apicurio Registry, two separate community repositories, a Quarkus LangChain4j integration proposal — with incompatible APIs, in the words of an Apicurio engineer from May 2026: “the interoperability window is closing — the longer we wait, the harder it gets to align.”

Every day an implementation ships, its assumptions about ownership, revocation and dispute handling harden a little more. And the economics point one way: a listing that costs nothing to create is a claim that costs nothing to dispute. Squatting is not a bug in any of these systems — it is the default behavior of any namespace where claims are free.

3. Three lessons from our own name wars

We are not writing about this from the balcony. In September 2026 we published the Resonant Time Transfer Protocol reference implementation across three ecosystems, and every one of them taught the same law differently. First: the bare name is already taken. The npm package rttp has been held since 2017 by an unrelated small HTTP library — seven years of silence, still a legal occupant. First come, first served is the default constitution of every namespace on the internet.

Second: where squatting is banned, publishing is the only way to hold. npm’s policies forbid empty placeholder packages — so we did not petition or wait; we published a real package with a real conformance suite (@aicent/rttp), and PyPI the same day (rttp), and crates.io earlier still. The name is held by the artifact, not by the request.

Third: verify-then-list changes what a listing means. The IANA URI Schemes registry — where rttp and iqa now sit as provisional entries, CRI 27 and CRI 28 — lists a name only with a named change controller and an expert review path. And an independent agent index listed our MCP tool on its own initiative, then granted the Verified Publisher badge only after we proved the publisher address and repository write access. Each of these listings binds to something that was checked. That is the difference between a directory and a land office.

4. What a registry owes the name it lists

Strip the vendor details and any registry must answer three questions. Who lists? — ownership proven at write time, not asserted in a profile. Who vouches? — a badge is a claim about a claim; a verifiable attestation is a claim you can recheck yourself. How does it expire? — a name without a staleness story is a name that outlives its truth.

Then the honest twist, from the addressing side of our stack: the Resonant Time Transfer Protocol (RTTP) needs no registry at all for its addresses — an address is derived from a canonical authority, so there is nothing to squat and nothing to grant; the WHERE question has no landlord. What still needs an answer is WHETHER the thing at the address stands — and that is Identity Quality Assurance (IQA): a sealed envelope, a named vouching subject, a 120-second lease, verifiable offline. Registries are useful as indexes; the vacuum closes when what they list binds to what was verified, not to who clicked first.

Sources　[1] Official MCP Registry — registry.modelcontextprotocol.io; modelcontextprotocol/registry README, quoted verbatim (“The registry validates namespace ownership when publishing”; namespace examples; API freeze v0.1, 2025-10-24), accessed 2026-10-18　[2] a2a-registry.org — claim/verification mechanics quoted from the site’s own pages (“Spot yours unclaimed? Sign in to claim it.”; DNS TXT / GitHub verification paths; +75 agents in 7 days), accessed 2026-10-18　[3] Forkast, “Three Agent Specs Promised for October Remain Unshipped...” (2026-10-10) — five implementations and the Apicurio May 2026 quotation, verbatim from the archived original (fetched 2026-10-18)　[4] First-hand registry experience, dated: PyPI rttp and npm @aicent/rttp published 2026-09-17; crates.io rttp held earlier; IANA CRI 27 (2026-09-22) and CRI 28 (2026-10-13); independent index listing and publisher verification, 2026-10　[5] Editorial note: individuals are cited by affiliation only; the 2017 npm occupant of rttp is described, not named.

One line: a registry is a promise that the name means what it says — and the vacuum gets filled either by whoever verifies first, or by whoever clicks first.