Who Ships Your MCP Servers? 37,220 Registry Entries Traced to Their Source
The official MCP registry is where coding agents and enterprise allowlists increasingly look up servers. By design it does not judge them: it leaves security checks to package registries and says downstream aggregators can add ratings and curation. So before asking whether an MCP server is safe, we asked something simpler: who ships it, and can you see what they ship?
We pulled every active server from the registry on 30 September 2026 (37,220 at their latest versions) and traced each one to its source. For the checks that need network lookups we drew a random sample of 1,000 servers and report each share with a 95% confidence interval.
One in five is just a URL
8,251 servers (22.2%) declare a remote endpoint and nothing else: no source repository and no package. You send that URL your prompts, your context and often a credential, and there is nothing to read.
That share is growing faster than the registry. On 6 August 2026 the registry had 20,095 active servers, 3,380 of them remote-only (16.8%). Eight weeks later it has 1.85× the servers and 2.4× the remote-only ones.
Many remote-only servers are hosted products from companies you may already trust, and being hosted isn't a flaw. It does mean your trust rests entirely on who runs the endpoint, because there is no code to inspect and no release to pin.
Another one in five points at a repository that isn't there
Most servers do declare a GitHub repository. In our sample, 761 servers declared one, covering 650 distinct repositories. 120 of those repositories (18.5%, 95% CI 15.7–21.6%) do not exist publicly. We confirmed a random 25 of them by hand: every one returned 404. Some are private, some deleted, some mistyped; from the outside you can't tell which.
About half of the servers behind those missing repositories don't publish a package either. Add them to the remote-only ones and 30.2% of registry servers (95% CI 27.4–33.1%) have no public source code you could read:
| Where the code is (sample of 1,000) | Servers | Share |
|---|---|---|
| A public GitHub repository | 610 | 61.0% |
| Remote URL only, or nothing declared | 224 | 22.4% |
| Declared repository missing, no package | 78 | 7.8% |
| Declared repository missing, has a package | 73 | 7.3% |
| Package only, no repository declared | 15 | 1.5% |
Who is the publisher?
The registry checks one thing about identity, and it's a useful one. A server named io.github.alice/… can only be published by someone logged in to GitHub as alice or as a member of that organisation. That lets you compare the account that published the entry with the account that owns the repository it points to.
Of the 22,388 GitHub-namespaced entries that declare a repository, 97.3% are published by the repository's owner. The other 604 (2.7%) point at a repository owned by a different account. We checked all 604:
| Publisher differs from the repository owner | Entries |
|---|---|
| The publisher has commits in the repository (usually a maintainer publishing for an organisation) | 381 |
| The declared repository doesn't exist publicly | 144 |
| The publisher has no commits in the repository | 53 |
| The repository is a fork, or was renamed to the publisher | 11 |
| Couldn't be checked (the publishing account no longer resolves) | 15 |
Those 53 are a small number, and many are probably benign: a company's second GitHub organisation, or a developer-relations account that publishes but doesn't commit. They are still the shape that impersonation takes. The 2025 postmark-mcp incident was a server published by someone other than the vendor it named, and it later shipped an update that copied every email to the attacker. A registry entry whose publisher has never touched the code it points to deserves a second look before you connect.
Packages mostly link home; few prove it
For servers that ship an npm or PyPI package, we checked the package against the repository the registry entry declares. Across the 357 packages in the sample:
- 84.9% (95% CI 80.8–88.2%) declare a source link that points back to that repository.
- 9.0% declare no source link at all, and 1.1% point at a different repository.
- Only 28.0% (95% CI 23.6–32.9%) carry a build attestation (npm provenance or PyPI's PEP 740). An attestation is the only one of these checks that proves the published files were built from that repository, rather than merely claiming it.
Active, but unaudited
The repositories that do exist look well maintained: 81.5% were pushed in the last 90 days and only 2.1% are archived. What they lack is independent evidence. Only 6 of 650 declared repositories (0.9%, 95% CI 0.4–2.0%) have a result in the public OpenSSF Scorecard API, the standard automated check of branch protection, code review, pinned dependencies and release signing. For almost every MCP server, nobody has published that check.
That gap is why HVTracker runs Scorecard itself for every project it scores, and why each profile page now opens with a chain of custody: from the source repository, through how changes are reviewed and released, to the package you install.
What provenance does and doesn't tell you
Provenance answers who ships the code and whether that changed. It doesn't tell you what the code does. An academic study of 1,899 MCP servers found that healthy project metrics coexisted with vulnerabilities and tool poisoning, and the incident record agrees: many of the most-reported MCP CVEs of 2025–26 were ordinary bugs in well-maintained projects. Content scanners have the opposite profile. They read tool descriptions and code, but a study of 64,611 servers found existing scanners flagged nearly 97% of them as risky, with fewer than half of sampled alerts being real.
The two complement each other. Provenance catches the fastest-growing kind of incident: malicious publishers, impersonation and typosquats, as in postmark-mcp and the ClawHavoc skills campaign. Content scanning catches injected instructions and risky tools in code that is otherwise legitimate. Use both, and treat neither as a verdict on its own.
What to do with this
If you install MCP servers:
- Prefer entries with a live public repository and a package that is attested and links back to it. Pin the version.
- Check that the registry namespace (
io.github.<account>) matches the repository owner. If it doesn't, check that the publisher actually works on the code. - For a remote-only server, decide whether you trust the operator as you would any SaaS vendor that receives your data, because that is what it is.
- Look up any server before you connect: hvtracker.net/verify, or the
verify_mcp_servertool on HVTracker's MCP server.
If you publish one: point the registry's repository field at a public repository, publish with npm provenance or PyPI trusted publishing so your packages are attested, and run the Scorecard action with publish_results: true so anyone can check it.
Method and limits
- Frame: every server with status
activein the official registry API (/v0/servers?version=latest), pulled 30 September 2026: 37,220 servers. Source type and namespace ownership are counted across the whole frame. - Sample: 1,000 servers drawn uniformly at random (seed 20260930). Package, repository and Scorecard checks run on the sample. Intervals are 95% Wilson intervals.
- Checks: package metadata from the npm and PyPI registries (source link, attestation on the latest release), repository existence, archive status and last push from the GitHub API, and Scorecard results from
api.scorecard.dev. A package link that names a renamed repository counts as a match. The 604 namespace mismatches were each checked for forks, renames and the publisher's commits on the default branch. - Limits: registry entries are self-declared. A missing repository may be a legitimate private one. Having commits shows involvement, not authorisation, and having none shows neither. The registry grows by hundreds of servers a day, so these are point-in-time figures.
Data: sample.csv (one row per sampled server) and metrics.json (every figure above, with counts and intervals), under CC BY 4.0. The study is reproducible with mcp_registry_pull.py and provenance_study.py.
Check a server before you connect it
Paste a repository, package or MCP server and get its provenance, evidence coverage and a signed verdict.
Verify an MCP server