MCP Server Landscape 2026: Verified Dataset + Report
14,973 MCP server rows joined across public sources and deduplicated, with published coverage for the licence, maintenance, auth, install and liveness fields. 5,340 canonical rows verified alive on 2026-08-25.
Is my server in there?
Search the public identity and liveness view. No API key needed; 5 results max.
Not in there? Submit it
Send a public HTTPS endpoint or GitHub, GitLab, or Bitbucket repository. Endpoints are probed now; admission uses the same automated gates at the next scan.
Checkout and delivery are handled by Gumroad. API docs, no key needed: /v1/schema.
| column | what it is |
|---|---|
| dedup_group | Forks, mirrors and package coordinates clustered across sources, not just within one. This is the product's main differentiator and it ships as a field. |
| commercial_use_ok | A mechanical mapping from the declared SPDX identifier, not the licence string itself, and not legal advice. Null where it depends on how you deploy, never a guess. |
| alive | Computed from the applicable evidence: repository reachability and recency where a repo exists, package installability where a package exists, endpoint reachability for declared remote endpoints. Checked on 2026-08-25, 5,340 canonical rows passed. |
| maintenance_status | Active, slow, dormant or abandoned, from the real last-commit date. Unknown on 50.2%, and the page says so. |
| license_spdx | Resolved SPDX identifier on 51.0% of rows. Never guessed from free text. |
| auth_required | None, API key or OAuth, read from declarative config only. |
| install_method | npx, uvx, docker, remote or manual, by a published rule set. |
| category | 22-term controlled taxonomy. 20.7% land on other, because 48% of the registry exposes only a name. |
| tool_count · tool_names | 56,395 tool names, statically extracted, on 66.8% of alive servers. Null means not declared, never zero. |
Coverage is not uniform, and here is where.
repo_url is present on 65.6% of rows. Every GitHub-derived field (stars, archived, last_commit_at) sits at 49.0%, because 2,476 rows declare a repository GitHub answers with a 404 and 5,154 are remote-only servers that never had one. license_spdx resolves on 51.0%. Filter on last_commit_at without knowing that and you drop half the file. The full per-field table ships in SCHEMA.md and in the report.
20.7% of rows are category = other, on purpose.
About 48% of the public MCP registry exposes a name and nothing else: no tools, no topics, no package. On those rows we refuse to guess. Three earlier builds reported 7.5%, 8.6% and 21%, and each was less true than the one after it. The number is a property of the registry, not of the pipeline.
Verified means alive and canonical. Nothing else.
Alive at the snapshot date, and not a duplicate of another row. It is not a security audit, no dependency was scanned, and no tool was called. If you need a security assessment, this is not one.
The free sample is a product sample.
Drawn from verified live servers, deliberately the good end of the corpus. The full set includes the honest tail, documented per field.
description is the only generated field.
A 40-row hand check found 2 imprecise values, which is an honest upper bound of about 5%. Wide interval, single sample, stated anyway. Every other field is deterministic and its rule is published.
How often is it updated?
The €4 download is a one-time snapshot dated 2026-08-25 and does not update. The API is re-checked continuously: liveness daily, updates every 3 days, MCP handshake daily. Freshness is promised weekly and served about every 3 days; every response carries an X-Snapshot-Date header so you always know how old the row is.
What licence do I get?
Non-exclusive commercial use of the compilation, one organisation, any number of people. Query it, transform it, merge it, build on it, quote up to 50 rows with attribution. Reselling the files as a dataset, directory or API is not included. The software in the rows is not ours and is not licensed to you by us, which is exactly why license_spdx and commercial_use_ok are columns. Full terms in LICENSE-DATA.md, which ships in the pack.
Is commercial_use_ok legal advice?
No. It is a mechanical mapping of a declared SPDX identifier against a published allow and deny list, and the list is in SCHEMA.md so you can check it. Where the answer depends on how you deploy, the column is null rather than a guess.
Final sale?
Yes. All sales are final on this digital download. Once you download the archive, the files cannot be recalled and there is no reliable way for us to verify deletion. Download the free sample first.
How many of the 14,973 are actually usable?
5,340 are verified: alive at the snapshot date and canonical. 5,226 also carry a resolved licence and a real maintenance date. The rest are in the file because knowing what exists and cannot be checked is also an answer, and alive = false is kept separate from alive = null.