Six entries in servers/ have a source.project that returns 404. The published images are fine —
I checked all six on Docker Hub and every one returns 200 with a live tag — so this is not a broken
server, it is a broken provenance link.
Flagging it because the source: block carries a pinned commit, which suggests it exists so
someone can verify what went into the image. For these six that check cannot be performed from the
catalogue.
| server |
source.project |
image |
Hub |
cdata-connectcloud |
CDataSoftware/connectcloud-mcp-server |
mcp/cdata-connectcloud |
200 |
pica |
picahq/mcp |
mcp/pica-mcp-server |
200 |
pulumi |
pulumi/mcp-server |
mcp/pulumi |
200 |
sqlite-mcp-server |
neverinfamous/sqlite-mcp-server |
mcp/sqlite-mcp-server |
200 |
tembo |
tembo-io/mcp-server-tembo |
mcp/tembo |
200 |
triplewhale |
Triple-Whale/mcp-server-triplewhale |
mcp/triplewhale |
200 |
The honest limit, and it is the important part of this issue: from outside, a private repo and a
deleted repo return the identical 404. I cannot tell those apart and I am not claiming to. Every one
of these owners still exists as a GitHub account, so the likeliest explanation for most of them is a
repo that went private or was renamed, not one that vanished. Two of the owners have visibly related
public repos under new names (CDataSoftware/cdata-jdbc-mcp-server, neverinfamous/postgres-mcp),
which looks like ordinary renaming rather than anything alarming.
So: six rows to look at, not six problems. If the answer is "those went private, working as
intended," that is a completely reasonable answer and this issue can close.
Reproduction, if it is useful — no clone of anything of mine, no account, writes nothing:
GITHUB_TOKEN=<a pat> npx -y -p github:siliroid/unreached unreached-registry ./path/to/mcp-registry
It walks the tree for github.com/<owner>/<repo> references, HEADs each one, and splits the 404s
into owner also gone vs owner alive — because those are very different stories and conflating
them turns a finding into an accusation. It also holds out rate-limited probes as UNKNOWN rather
than counting them dead, and ignores references that appear only in prose (CONTRIBUTING.md etc),
since documentation examples are not catalogue entries.
Full disclosure on that tool: pointed at this repo it initially produced 9 findings, and 3 of
them were mine, not yours — user-attachments/assets (a GitHub CDN path, not a namespace) and two
placeholder orgs from CONTRIBUTING.md and docs/configuration.md. Fixed before filing. Its
first-contact false-positive rate on a registry it had never seen was 33%, which is the number
I would want if I were reading someone else's audit.
Six entries in
servers/have asource.projectthat returns 404. The published images are fine —I checked all six on Docker Hub and every one returns 200 with a live tag — so this is not a broken
server, it is a broken provenance link.
Flagging it because the
source:block carries a pinnedcommit, which suggests it exists sosomeone can verify what went into the image. For these six that check cannot be performed from the
catalogue.
cdata-connectcloudCDataSoftware/connectcloud-mcp-servermcp/cdata-connectcloudpicapicahq/mcpmcp/pica-mcp-serverpulumipulumi/mcp-servermcp/pulumisqlite-mcp-serverneverinfamous/sqlite-mcp-servermcp/sqlite-mcp-servertembotembo-io/mcp-server-tembomcp/tembotriplewhaleTriple-Whale/mcp-server-triplewhalemcp/triplewhaleThe honest limit, and it is the important part of this issue: from outside, a private repo and a
deleted repo return the identical 404. I cannot tell those apart and I am not claiming to. Every one
of these owners still exists as a GitHub account, so the likeliest explanation for most of them is a
repo that went private or was renamed, not one that vanished. Two of the owners have visibly related
public repos under new names (
CDataSoftware/cdata-jdbc-mcp-server,neverinfamous/postgres-mcp),which looks like ordinary renaming rather than anything alarming.
So: six rows to look at, not six problems. If the answer is "those went private, working as
intended," that is a completely reasonable answer and this issue can close.
Reproduction, if it is useful — no clone of anything of mine, no account, writes nothing:
It walks the tree for
github.com/<owner>/<repo>references, HEADs each one, and splits the 404sinto owner also gone vs owner alive — because those are very different stories and conflating
them turns a finding into an accusation. It also holds out rate-limited probes as UNKNOWN rather
than counting them dead, and ignores references that appear only in prose (
CONTRIBUTING.mdetc),since documentation examples are not catalogue entries.
Full disclosure on that tool: pointed at this repo it initially produced 9 findings, and 3 of
them were mine, not yours —
user-attachments/assets(a GitHub CDN path, not a namespace) and twoplaceholder orgs from
CONTRIBUTING.mdanddocs/configuration.md. Fixed before filing. Itsfirst-contact false-positive rate on a registry it had never seen was 33%, which is the number
I would want if I were reading someone else's audit.