How we rank
Lists on agentbuffs are ordered by a score from 0 to 100. This page explains what goes into the score and what the labels mean. Every number comes directly from the site code.
The score
Every item has a quality score. The default sort, Recommended, lists items by this score, highest first. Ties break by stars, then by name. The formula is at version 2. An item page shows its score, score components, and this version.
| Part | Weight | What it measures |
|---|---|---|
| Popularity | 27% | Stars (75%) and forks (25%) on a log scale, compared with items of the same type. MCP servers are compared with MCP servers, and skills with skills. |
| Maintenance | 22.5% | The last push (80%) and the latest release (20%). A push holds full value for 30 days, then halves every 180 days. A release holds full value for 365 days, then halves every 365 days. |
| Trust | 18% | An official maker (40%), a verified domain namespace (15%), an open license (25%) and an active registry entry (20%). |
| Completeness | 13.5% | An install snippet (35%), a description of at least 40 characters (25%) and the share of known key facts, such as pricing, auth and hosting (40%). |
| Community | 9% | The average user rating, once 3 people rate the item. Neutral before that. |
| Agent reports | 10% | Install success (60%), ease (20%) and satisfaction (20%), as reported by AI agents. Neutral below 10 networks. |
Missing data is neutral, scored at 50 out of 100. It neither helps nor hurts an item.
Penalties multiply the whole score. Penalties stack.
| Penalty | Multiplier | When it applies |
|---|---|---|
| Archived | ×0.25 | The repository is archived. |
| Deprecated | ×0.4 | The registry marks the item as deprecated. |
| No push in two years | ×0.6 | No push for more than 730 days. |
| No license | ×0.75 | No license file was found. |
| Nothing to install from | ×0.5 | No package, remote, repository or plugin exists to install from. |
Stars belong to a repository, not an item. Items from the same repository share its stars and may score close together. In the default skills view, the third and subsequent skills from one repository move after all other skills. They are not removed; they appear on later pages. Top lists on the home page, category pages, and client pages show at most 2 items per repository. Best for picks apply the same limit. The MCP servers list, and any sorted or searched list, stays in strict score order.
Official
An item is Official when published by the company behind the service. Any one of these qualifies an item:
- An editor set it by hand.
- The punkpeye/awesome-mcp-servers list marks it as official.
- An official skills list marks it: the anthropics/skills repository, or an "Official skills by" group in VoltAgent/awesome-agent-skills.
- It is a reference server in modelcontextprotocol/servers.
- The MCP Registry verified the publisher domain as the vendor domain, or the repository owner matches the vendor name.
Official indicates publisher identity. It does not mean agentbuffs reviewed the code or that the item is safe. Curated lists and name matches can be wrong; an editor can correct them. Community means any other item. It is not inferior, only published by someone else. "Maker: Official" and "Maker: Community" show this distinction.
Official counts under Trust in the score. Compared to a community item, it adds 7.2 points out of 100.
Featured
Featured is an editor's pick. A person selects each item and writes a reason of up to 120 characters. No algorithm selects it, and importers never alter it. A pick shows only while the item remains public and active.
Featured does not alter the score or the Recommended sort. Picks appear on the home page (up to 8), on the Featured page, and pinned atop their category page (up to 4).
Best for
The sort options Best for first-timers, developers, teams, and privacy each run on a separate 0 to 100 score. Each score weights item signals and applies the standard penalties. It reads only item attributes, never user data. Missing signals count as neutral.
| Sort | Favours | Signals and weights |
|---|---|---|
| Best for first-timers | Easy to set up: runs remotely or installs with one command, simple sign-in, from the maker. | Easy install 30%, Easy sign-in 20%, From the maker 15%, Ease of setup rating 10%, Maintenance 10%, Completeness 15% |
| Best for developers | Runs locally or ships scripts, has an open license, is actively maintained and widely used. | Runs locally or ships scripts 25%, Open license 20%, Maintenance 20%, Popularity 25%, Completeness 10% |
| Best for teams | From the maker, signs in with OAuth, is maintained, has a permissive license and can be read-only. | From the maker 25%, OAuth sign-in 15%, Maintenance 20%, Permissive license 20%, Read-only access 10%, Trust 10% |
| Best for privacy | Runs on your machine, needs no cloud service, is open source and needs no network. | Runs on your machine 30%, No cloud service 20%, Open license 25%, Maintenance 15%, Completeness 10% |
Above the default list on page 1, with no filter, search or sort applied, a strip names the top item for each of the four Best for options. A pick requires a score of at least 60, a list of at least 6 scored items, and at least 2 reasons displayed as "Good fit for" chips. An item wins at most one pick. Archived items never win.
Maintenance
This label tracks repository activity by last push date.
| Label | Rule |
|---|---|
| Active | Pushed within the last 90 days. |
| Slow | Last pushed between 90 and 365 days ago. |
| Stale | Not pushed for more than 365 days. |
| Archived | The repository is archived. |
| Unknown | No push date is known, and the repository is not archived. |
The label is distinct from the Maintenance score component. The score drops smoothly over time. An item without a push for over 730 days also incurs the stale penalty.
Ratings and agent reports
User ratings and AI agent reports appear separately. They never combine.
People: members with confirmed emails rate items from 1 to 5 across four areas: usefulness, ease of setup, documentation, and reliability. The average displays once 3 people rate the item. Below that threshold, the page shows "Not enough ratings yet" and the score treats Community ratings as neutral.
AI agents: agents report via the public API whether installation succeeded and whether the tool worked. People do not submit these. Reports group by network (IP address blocks), so one operator counts once. Data displays once reported by 5 networks.
Reports affect the score only at 10 or more networks. The algorithm shrinks values toward neutral, as if 10 additional networks reported neutral results. At weight 10%, agent reports shift the score by at most 5 points.
Where the data comes from
Facts about each item come from public registries, curated lists and GitHub. The sources are listed on the About page.