State of Agent Add-ons

As of 31 October 2026 our directory holds 1,233 add-ons: 438 plugins, 426 MCP servers, 183 subagents, 95 workflows, 62 skills, 18 hooks, 8 rules files and 3 slash commands. This report counts what we list, not the whole market. All entries are still drafts until each passes our publishing gate.

What this report is

These are counts from our own directory data, taken on 31 October 2026 from the local working copy of the database, read-only. They describe our collection. They do not describe how many add-ons exist, and they do not say which are good.

Two limits to hold in mind throughout:

1. Coverage follows our sources. Most entries came from a few public lists, set out below. Add-ons that are not on those lists are not here.
2. Nothing is published yet. Every entry is a draft until an editor adds a note and the entry passes the publishing gate. Counts will change as entries pass or are removed.

The headline numbers

Add-ons by type

Shares are rounded. Plugins and MCP servers make up about seven in ten of what we hold. See the hubs for MCP servers, skills and subagents.

The small counts for skills and hooks do not mean few exist. Skills often live inside larger repositories and plugin bundles, so they are not always listed as separate entries.

Where the entries came from

This is the reason the report is about our collection. Five of the six sources are built around Claude Code or are MCP-specific. Add-ons made for other agents are underrepresented until we add sources for them. Every entry stores its source name and link, and the page shows it.

Add-ons by category

882 of the 1,233 add-ons have a category. The rest are uncategorised so far.

Web search and scraping leads, followed by observability and security. We have not analysed why. One likely factor is that many vendors publish an MCP server for their own product, and the official registry is one of our sources.

License

About 86% of the add-ons we list carry an open-source license class. Missing license data is a gap in our records, not a finding about the add-on.

Official versus community

131 add-ons are flagged official, meaning published by the vendor of the product they connect to. Another 697 are not, and 405 have no flag yet.

Of the 131, 59 are official MCP servers that also carry an open-source license class.

How MCP servers connect

MCP servers can run on your machine over stdio, or remotely over HTTP. Of the 426 MCP servers, 391 have transport data.

Local stdio servers are still the larger group among those with data. Remote servers matter for teams because there is nothing to install. We have no data on how the split changes over time. See what are MCP servers.

Freshness

1,215 add-ons have a last-commit date from their repositories. Of those, 1,160 had a commit in the last 90 days, and 1 is marked archived. Our entries are recent partly because the sources we draw on are recent lists. An add-on that stopped updating a year ago is less likely to be on them.

Per entry, we also record a last-verified date once an editor has checked the facts. Nine entries have one so far. That is the number that matters for the publishing gate, which needs a check within 90 days.

Languages and packaging

Where the primary language is recorded, the most common are Python (405 entries), TypeScript (237), Shell (214), JavaScript (105), Go (58) and Rust (41). For package registry, npm (146) and PyPI (86) lead, with container images (29) and NuGet (7) behind. Many entries record neither.

What to take from this

• For choosing: the add-on space is wide and shallow. Hundreds of entries in a category does not mean hundreds worth installing.
• For building: the gaps are in the small categories at the bottom of the table and in non-Claude agents.
• For reading any directory: ask where its entries came from.

For which agents can use which add-on formats, see the compatibility matrix. To add something we missed, use the submit form.

Method

• Counts were run with read-only queries against the local D1 database on 31 October 2026.
• Types, categories, flags and transports are as recorded on each entry, not re-checked by hand for this report.
• Percentages are rounded to the nearest whole number.
• Entries are drafts. We will refresh this report after publishing starts. The rules are in the methodology, and mistakes can be reported through the corrections policy.