Agent Plugins Statistics

As of 2026-09-21, there are 2765 verified Agent Plugins across 1355 GitHub repositories, shipping 11,678 Agent Skills and 1260 MCP server entries. Source: Agent Plugins Directory, https://agentpluginsdirectory.com/stats

As of 2026-09-21, there are [2765 verified Agent Plugins](https://agentpluginsdirectory.com/stats) across 1355 GitHub repositories (Agent Plugins Directory).

As of 2026-09-21, 2,765 verified Agent Plugins exist across 1,355 public GitHub repositories. agentpluginsdirectory.com fetched 4,840 conformant plugin.json manifests from 1,480 repositories, then set aside 751 from two bulk collections, 647 test fixtures and 677 duplicate publications of a plugin already counted. Those 2,765 plugins ship 11,678 Agent Skills and 1,260 MCP server entries. The count changed by +59 (+2%) since the previous crawl on 2026-09-20.

No official registry publishes this number. The Agent Plugins 1.0.0 specification ships none.

How many Agent Plugins are there?

Measure	Count
Distinct verified plugins	2,765
Conformant manifests found	4,840
GitHub repositories with at least one manifest	1,480
Repositories behind the 2,765	1,355
Agent Skills shipped by those plugins	11,678
MCP server entries shipped by those plugins	1,260
Manifests still pinning the pre-rename schema	6

The Agent Plugins 1.0.0 specification was published on 2026-08-06. Every count on this page reflects the state of GitHub 46 days later, on 2026-09-21.

agentpluginsdirectory.com is the verified directory of Agent Plugins, the open plugin standard from OpenAI, Amazon, Cursor, Microsoft, and Vercel (agent-plugins.org) supported by ChatGPT, Codex, Cursor, GitHub Copilot, VS Code, and Kiro. Every listing is verified by fetching its plugin.json manifest and checking it against the official 1.0.0 schema.

What changed since the previous crawl

We run the same crawl, with the same exclusion rules, every day. The table compares the two most recent runs.

Measure	2026-09-20	2026-09-21	Change
Distinct verified plugins	2,706	2,765	+59 (+2%)
Conformant manifests	4,770	4,840	+70 (+1%)
Repositories with a manifest	1,460	1,480	+20 (+1%)
Agent Skills	11,558	11,678	+120 (+1%)
MCP server entries	1,247	1,260	+13 (+1%)

619 of the 1,355 repositories behind the count were created on or after the specification's launch date, 46% of the total. The rest are existing projects that added a manifest.

For scale: the first census on 2026-08-07 found 257 verified plugins, and the second on 2026-08-08 found 488, a 90% rise in the ecosystem's second day. The full day-by-day series accumulates in the published dataset.

How we counted

Search engines return fuzzy matches. A GitHub code search for the schema string surfaces client source code, blog drafts and README snippets alongside real manifests, so a raw result count overstates the ecosystem by a wide margin.

Our method, in order:

1. Query the GitHub code search index for the exact canonical $schema string with filename:plugin.json, splitting the query with -repo: exclusions to break the 1,000-result API cap.
2. Add a repository search for agent-plugins created after the spec launch, because GitHub's code index lags several days behind new repositories.
3. Add a separate query for the legacy open-plugins.com schema identifier.
4. Enumerate the repositories of the launch-partner organizations directly.
5. Fetch the raw plugin.json for every candidate and validate it against the official 1.0.0 schema stored at data/schemas/plugin.schema.json.
6. Discard everything that fails.
7. Set aside the manifests that pass validation but are not a distinct installable package: bulk auto-generated collections, extension-namespace stubs, conformance test fixtures, and repeat publications of a plugin already counted.

Step 5 is what separates this count from an estimate. The official schema requires both $schema and name, pins $schema to a constant value, and sets additionalProperties: false, so a file either validates or it does not. The specification agrees: a manifest missing a required field is invalid, and per §5 "Clients MUST reject the plugin and MUST NOT discover or execute any of its components."

Step 7 is the one that moves the headline. Validating and being installable are different tests, and 2,075 manifests pass the first while failing the second. Conformance suites ship manifests that are designed to be broken, with names like invalid-mcp-json, broken-mcp, Has--Double and from-the-future, and a wrapper manifest whose mcp.json is deliberately malformed still has a schema-valid plugin.json beside it. A directory that counts invalid-mcp-json as a verified plugin has verified nothing. We filter on the manifest path rather than the publisher, because the repositories that ship fixtures also ship real work: in the 2026-08-08 census, stbenjam/skillsaw contributed 17 fixtures and 7 genuine plugins, and only the 17 were removed.

Why one plugin can be published several times

A publisher builds a plugin in a monorepo and then republishes it from standalone, per-client repositories, so the identical package appears at three GitHub addresses. remotion v4.0.507 shipped from remotion-dev/remotion, remotion-dev/codex-plugin and remotion-dev/cursor-plugin at once; Sentry, LangWatch and others do the equivalent. Every copy declares the same repository upstream in its own manifest, which is what ties them together, and that is what we group on.

Ranking those copies by popularity would be the obvious move and the wrong one. For both remotion and sentry the most-starred copy is the one declaring zero components, because its skills do not sit at the specification's fixed location relative to that manifest — a conformant client loading it finds nothing. We keep the copy with the most components, preferring a manifest at the repository root, and fall back to stars only to break a tie.

A separate correction is folded into the same rule: GitHub repository names are case-insensitive and resolve to one repository, but our crawl had been indexing each spelling it encountered, so mylokaye/foundations and mylokaye/Foundations — one repository, id 1326295889 — were contributing four plugins that do not exist. That was a defect on our side, not an editorial judgement, and it is corrected here rather than quietly.

Collapsed listings are not deleted. Their URLs permanently redirect to the surviving plugin, because they were live and indexed before this rule existed.

What the categories are, and what they are not

Every verified plugin is filed under exactly one of 13 categories. That grouping is the one number on this page that is not verified, and the distinction matters. Agent Plugins 1.0.0 defines no category field, and the manifest schema sets additionalProperties: false, so no publisher can declare one even if they want to. Every category here is our reading of what a plugin does, from its name, description, declared keywords, skill names and MCP server names.

Three rules keep that inference honest. A plugin is classified once and the answer is pinned permanently, so nothing silently changes category between crawls. Corrections are made by hand in an overrides file that always wins. And "Other" is a real answer rather than a dumping ground: a plugin whose manifest describes too little to place lands there instead of being guessed into a category that would mislead you.

Category	Plugins	Share
Other	1,043	38%
Agent Tooling	352	13%
Integrations & APIs	307	11%
Productivity & Workflow	194	7%
DevOps & Cloud	137	5%
Research & Search	132	5%
Code Review & Quality	121	4%
Security & Compliance	106	4%
Documentation & Writing	100	4%
Design & Frontend	93	3%
Databases & Data	81	3%
Creative & Media	64	2%
Testing	35	1%

Why 4,840 manifests becomes 2,765 plugins

Four exclusions account for 2,075 of the 4,840 manifests, or 43% of everything published. Counting them at face value would nearly triple the apparent size of the ecosystem.

Excluded	Manifests	Why it is set aside
withoneai/one-agent-plugin	651	Auto-generated integration wrappers, one per third-party app, published in a single commit by one publisher
github/awesome-copilot	100	Stub manifests that declare no components at the specification's fixed locations; skills are declared through the com.github.awesome-copilot extension namespace instead
Test fixtures, 74 repositories	647	Manifests under test/, fixtures/ and examples/ paths, most of them deliberately invalid packages written to exercise a validator's rejection path
Duplicate publications	677	The same plugin published more than once — from a monorepo and again from standalone per-client repositories, from two paths inside one repository, or from one repository indexed under two spellings of its name. Grouped by plugin name plus the repository upstream the manifest itself declares

4,840 conformant manifests
  - 651 withoneai/one-agent-plugin (auto-generated wrappers)
  - 100 github/awesome-copilot (extension-namespace stubs)
  - 647 conformance test fixtures (74 repositories)
  - 677 duplicate publications of a plugin already counted
  = 2,765 distinct verified Agent Plugins, across 1,355 repositories

The fixture population is dominated by tooling projects. In the 2026-08-08 census, five repositories supplied 106 of the 127 fixtures — HiAi-gg/agent-plugins-doctor (33), kuralle/kuralle-agents (26), stbenjam/skillsaw (17), Toasterson/agent-plugin-rs (15) and Toasterson/agent-plugin-ts (15) — and every one of them is a linter or conformance harness, which is a reasonable signal about where tooling effort is going. Read the full derivation in how we counted the first verified plugins.

Skills versus MCP servers

Version 1.0.0 standardizes two component types: Agent Skills discovered under skills/, and MCP servers declared in a root mcp.json. Skills dominate.

Composition	Plugins	Share
Ship at least one skill	2,155	78%
Ship at least one MCP server	1,083	39%
Ship both	811	29%
Ship neither (metadata only)	338	12%

Two readings follow from that split. Agent Skills are markdown files with frontmatter, so authors can port an existing skills folder into a plugin without touching runtime code. MCP servers need a process, a transport and often credentials, which the 1.0.0 specification does not yet address (its FUTURE_CONSIDERATIONS.md defers secret handling to a later version).

Splitting the composition further: 1,344 plugins ship skills alone, 272 ship MCP servers alone, and 811 ship both. Only 338 plugins ship neither component. Most of those are placeholders published to claim a name, or packages whose manifest sits beside components the spec does not look for. MCP server entries went from 1,247 to 1,260 between 2026-09-20 and 2026-09-21.

License distribution

2,075 of the 2,765 plugins declare a license. 690 declare none.

License	Plugins
MIT	1,367
(none declared)	690
Apache-2.0	479
Proprietary	93
(MIT AND CC-BY-SA-4.0)	41
UNLICENSED	14
GPL-3.0-or-later	10
AGPL-3.0-or-later	10
AGPL-3.0-only	5
AGPL-3.0	5
CC-BY-4.0	5
ISC	4
LicenseRef-CoCo-Bundled-Plugin-Permission	4
CC-BY-SA-4.0	3
MIT-0	3
GPL-3.0	3
MPL-2.0	2
PolyForm-Noncommercial-1.0.0	2
Dual: CC BY 4.0 / MIT	2
Elastic-2.0	2
BSD-3-Clause	2
CC0-1.0	2
17 further identifiers, one plugin each	17

MIT and Apache-2.0 together cover 1,846 plugins, 67% of the total. The specification does not require license to be a valid SPDX identifier, and a handful of manifests exercise that latitude with values like Proprietary, UNLICENSED or LicenseRef-Candor-Proprietary, none of which a legal review can resolve to terms.

Field adoption

Only $schema and name are required. Optional fields fill in unevenly.

Field	Plugins declaring it	Share
description	2,736	99%
version	2,613	95%
keywords	2,239	81%
license	2,075	75%
extensions	940	34%

The extensions figure is the one to watch. That field is where client-specific data lives, and its low adoption so far suggests most publishers are writing one portable manifest rather than tailoring per client.

Top publishers by plugin count

Publisher	Plugins	Repos	Repo stars
rakthainet-DKH	96	1	0
JoAiHQ	79	1	0
mjcramerz	51	1	0
Agents-Store	51	1	0
reforma-ai	44	2	5
open-coder-ai	42	2	10
netresearch	41	41	577
insodimension	39	1	1
sickn33	37	1	46,689
SpillwaveSolutions	37	37	237
vellum-ai	36	1	1,293
reforma-dev	36	1	0

Plugin count and popularity diverge sharply. In the 2026-08-08 census, sickn33 published 58 plugins from one repository carrying 44,651 stars, while netresearch published 40 plugins across 40 separate repositories that totaled 468 stars between them. Three of the top five publishers shipped everything from a single repository.

This table is where excluding test fixtures shows most. In that same census, HiAi-gg ranked fifth with 16 plugins after 33 of its manifests were removed as fixtures, and Toasterson left the list entirely because all 30 of its manifests are conformance cases.

The 1,355 repositories belong to 1,106 distinct owners: 489 GitHub organization accounts and 617 individual accounts. On 2026-08-07, organizations led 53 to 49; a day later individuals had overtaken them. The second wave of publishing came from people, not vendors. Vendors wrapping their own commercial product, including AWS, Google Cloud, Netlify, Neon, Fastly, GoDaddy, Sentry, Resend, RevenueCat, Svix, Mobbin and Brandfetch, are listed on best Agent Plugins.

Most-starred repositories shipping a plugin

Star counts belong to the host repository, not the plugin, and most of these repositories earned their stars for the product rather than the manifest.

Repository	Stars	Plugins	Skills	MCP
koala73/worldmonitor	87,129	1	25	2
upstash/context7	62,267	1	1	1
ChromeDevTools/chrome-devtools-mcp	52,414	1	7	1
sickn33/agentic-awesome-skills	46,689	37	308	0
K-Dense-AI/scientific-agent-skills	45,870	1	166	0
saadeghi/daisyui	42,441	1	1	0
volcengine/OpenViking	38,289	1	3	1
huggingface/diffusers	34,555	1	4	0
feder-cr/AIHawk	31,614	1	1	1
pascalorg/editor	24,180	1	2	1
vectorize-io/hindsight	24,168	1	1	1
different-ai/openwork	23,682	1	1	1
google/skills	20,225	1	5	1
superset-sh/superset	14,433	1	13	0

Four repositories in this table were absent from the 2026-08-07 index: exa-labs/exa-mcp-server, langwatch/langwatch, DannyMac180/sol-advisor and AvdLee/Swift-Concurrency-Agent-Skill. Established projects with four-figure star counts are adding manifests, which is a different adoption signal from new repositories appearing. The full ranked list lives on best Agent Plugins.

Schema versions found in the wild

$schema value	Manifests
https://agent-plugins.org/schemas/1.0.0/plugin.schema.json	4,814
https://agent-plugins.org/schemas/2.0.0/plugin.schema.json	8
https://open-plugins.com/schemas/1.0.0/plugin.schema.json	6
https://agent-plugins.org/schemas/9.9.9/plugin.schema.json	4
https://agent-plugins.org/schemas/1.1.0/plugin.schema.json	3
http://agent-plugins.org/schemas/1.0.0/plugin.schema.json	2
https://agent-plugins.org/schemas/0.9.0/plugin.schema.json	1
https://agent-plugins.org/schemas/99.0.0/plugin.schema.json	1
https://agent-plugins.org/schemas/v1/plugin.schema.json	1

This table counts all 4,840 manifests, before the exclusions above. 6 manifests across 5 repositories still pin open-plugins.com, the identifier used before the project renamed itself on 2026-07-17, and those are real plugins that stay in the count. Manifests declaring schema versions that do not exist (0.9.0, 2.0.0, 9.9.9, 99.0.0) are deliberate negative test fixtures in validator projects; all of them sit under fixture paths, so all of them are excluded. See the consortium's own legacy manifests.

Caveats we are publishing on purpose

The fixture rule is a path rule, so it is blunt in both directions. A real plugin published under a directory named examples/ would be excluded, and a fixture parked at a repository root would be counted. We chose the rule that is reproducible from the dataset over one that needs a human to judge each manifest, and the excluded list ships with the data so the choice can be checked.

The 2,765 verified manifests resolve to 2,763 distinct (repository, name) pairs. Across the set, 2,418 plugin names are unique and 133 names are claimed by more than one repository. As of 2026-08-08, sentry was used by four unrelated publishers; context7, github, remotion and example-plugin by three each. The specification defines no namespace, so name collisions are legal and will get worse.

GitHub's code search index lags days behind new repositories, which means this count is a floor rather than a ceiling. The 90% jump between 2026-08-07 and 2026-08-08 was partly real growth and partly the index catching up on repositories that already existed. We cannot separate the two from outside GitHub, so we report the combined figure and say so.

How this number will change

Our crawler runs daily and commits the refreshed dataset, so the git history of all.json is a time series of the ecosystem from day two forward. Numbers on this page regenerate from that dataset on every run and carry the date they were verified. The series is only days old, and we will not extrapolate from it.

To check whether your own manifest would be counted, run it through the plugin.json validator. To publish one, start with the guide to building an Agent Plugin or browse the plugin directory.

Use this data

Every number on this page comes from one dataset, and that dataset is public. Fetch the nested source as JSON or the flat one-row-per-plugin distribution as CSV:

curl https://agentpluginsdirectory.com/census.json
curl -O https://agentpluginsdirectory.com/census.csv

It carries the summary counts, the full per-category breakdown, the exclusion rules as structured data, and one record per verified plugin with its manifest path, schema identifier, skills, MCP servers, license and publisher. It regenerates on every crawl, so it never disagrees with the prose above.

Released under CC BY 4.0 — use it in anything, including commercially, as long as you attribute it. Cite as:

agentpluginsdirectory.com verified index, 2026-09-21. https://agentpluginsdirectory.com/stats

If you publish an analysis built on this data, we would like to read it. There is no other census of this ecosystem to check it against, which is exactly why we publish the raw file alongside the conclusions rather than asking anyone to take the numbers on trust.

Verified on 2026-09-21 by agentpluginsdirectory.com. Every figure comes from fetching each plugin.json from GitHub and validating it against the official Agent Plugins 1.0.0 schema. Cite as: agentpluginsdirectory.com verified index, 2026-09-21.