Journal

Why making our GitHub membership public did not fix registry publishing

Publishing our browser tool to the Model Context Protocol Registry failed with a 403 that told us to make an organisation membership public. The membership was already public. The registry had stopped using that mechanism, and its own sign-in flow could never reach the organisation namespace.

Published 7 min read

Our browser tool is published as a package that other agent systems install, and it is listed in the Model Context Protocol Registry, a public directory where agent tools are catalogued and where the listing pins the exact package version a client will install. On October 13, 2026, we published version 1.10.0 to that registry by hand. Getting there took longer than the publish itself, because the registry's error message sent us to fix a setting that was already correct.

What happened

Automated release publishing worked for our packages and failed for the registry. Our release automation published the JavaScript packages and the Python package on each release. Behind the registry listing there was no automation, so it was updated by hand, and each hand publish met the same response: a 403 error advising us to make an organisation membership public, so the registry could confirm that we were entitled to publish under the organisation's name.

We checked the membership. It was public. It had been public before the error told us to change it. Repeating the publish returned the same 403 and the same advice, while the state the advice described was already in place.

RouteHow it signs inNamespace it can reach
Interactive registry loginA device-flow sign-in in a browser, using a token the registry's own flow requestsPersonal namespace only, because that token cannot read organisation membership
Release automation after the fixOpenID Connect sign-in from continuous integration, issued to the release job itselfOrganisation namespace, because that route can establish the caller's organisation role

The timeline

WhenWhat happened
Before October 13, 2026Release automation publishes the JavaScript and Python packages; the registry listing is updated by hand and returns a 403 advising a membership change that is already in place
October 13, 2026Version 1.10.0 is published to the registry by hand, as a stopgap
October 13, 2026, same dayThe cause is confirmed in the registry's own source, and the registry step is added to the release automation with an OpenID Connect sign-in, a version sync, and a rule that a failed registry step fails the release run

What was actually wrong

Who may publish under an organisation's name is decided by the registry, by asking whether the caller's token can read the caller's organisation membership role. That check needs a specific permission on the token, the permission to read organisation membership. The interactive sign-in uses a GitHub App token that cannot read those memberships. The registry documents that limitation.[1]

Without the permission, the registry quietly treats the caller as an individual. A publish under that token then lands in a personal namespace, or fails against the organisation namespace with the 403 we saw. What the error message described was an older version of this arrangement, in which making a membership public was the step that let the registry see it. Public visibility was no longer the mechanism. In fact the setting was correct and the token was the problem, and the message pointed at the setting.

We confirmed this in the registry's own source and against the live service on October 13, 2026. Reading the source settled a question that repeating the publish could not: the interactive route was incapable of reaching the organisation namespace, however carefully we followed the error's advice.

One limit on this account belongs in the open. We have no count of how many attempts we made before reading the source, and this post invents none. In short, the error advised a change, the change was already in place, and the source explained why the advice could never have worked.

What we changed

Our fix puts the registry step inside the release automation, on the path that can actually reach the organisation namespace. After the JavaScript package publish succeeds, the release job signs in to the registry with OpenID Connect, using an identity token issued to the job itself. That route needs no dedicated registry secret. The registry accepts the identity of the release job for publishing under the organisation name.[1][2]

Two details of the step matter. First, the automation syncs the registry file's two version fields, the server version and the package entry, from the package file, so no version number in the registry file is edited by hand again. Strict ordering applies as well: the registry step runs after the package publish, because the registry checks that the package version it is being told about already exists. A registry entry that points at a version nobody can install would trade one kind of drift for another.

Failing loudly is wired into the step as well. If the registry publish fails, the release run fails. Before this change, a release could reach the package registries and skip the directory entirely, and nothing in the run reported the gap. Directory installs, and the mirrors that read the registry, would sit one version behind with no signal anywhere.

After the fix, all three publishing destinations run in the release automation (Figure 1).

All three publishing destinations are automated after the fixDestinations automated, out of 3: npm, PyPI, and the Model Context Protocol Registry
Before the fix2
After the fix3

Before the fix, the registry listing was published by hand.

After the fix, the registry step runs in the release automation and a failure fails the run.

Automation landed on October 13, 2026, the same day as the hand publish it replaces, and its version sync was exercised locally against the registry's validator at the time. The public registry later recorded version 1.15.0 on October 25, 2026.[3] We also verified a successful registry publish in our release logs on November 20, 2026.

What it cost

Direct cost was time spent following advice that described a mechanism the service no longer used, plus one hand publish of version 1.10.0 as a stopgap on October 13, 2026. We kept no record that would support an hours figure, so this post gives none.

Standing cost, until the automation landed, was drift. A registry entry pins the exact version its clients install. Every release that reached the package registries while the registry listing stayed behind left directory users installing the older version, and left the mirrors that ingest the registry showing the older version as current. Drift of that kind produces no error anywhere. Who finds it is a user who installed the tool and got last month's release.

What we take from it

An error message is a claim about a system, written at the time its author last understood the mechanism. Services change their mechanisms. Messages that describe the old one stay in the code and keep sounding authoritative, because they arrive in the voice of the system itself. Resolving this incident cost little: compare the message against the observed state, and when the two disagree, read the source of the service making the claim.

For another team, the practical version is a short sequence. When an error tells you to change a setting, verify the setting first. When the setting is already as advised, stop repeating the action that produced the error and go to the source, the documentation of the current mechanism, or the service's own responses to a differently-shaped request. In our case a differently-shaped request, a sign-in route whose token could read organisation membership, was also the fix, and it belonged in automation from the start: a publish step a person has to remember, against a sign-in flow that cannot complete it, will drift. Automating the path that works, and making its failure fail the run, removed both problems at once.

References

[1] MCP Registry: Authentication

[2] MCP Registry: Publishing from GitHub Actions

[3] Official MCP Registry: Rove 1.15.0