How we grade

Every plugin on zoteroplugins.org gets a Zotero Plugins grade, from A+ to C. We download its latest release and read its code, and the single worst finding sets the grade, so a plugin can't make up for a serious problem by doing well elsewhere.

The ladder

A+Recommended, tested in ZoteroPassed the code checks, and behaved as its card describes installed in a live Zotero · 178 pluginsARecommendedEvery check of the code came out low · 630 pluginsBUse with careRead why before installing · 673 pluginsCNot recommendedWe found serious problems · 76 plugins–Not graded yetWe couldn't check it yet · 16 plugins

An A doesn't mean we found nothing. Small findings, such as storing an API key in Zotero's settings, are rated low and listed on the card; an A means none of them is more than that. A B means at least one finding is worth reading before you install. A C means we found something serious, such as code deliberately scrambled so nobody can check it.

When a plugin has network code we couldn't trace to an address, it shows as not graded until we've checked further. If we ran it in Zotero and it contacted nothing, that counts as low concern.

For the short answer, with the problems we find most often and how to check a plugin yourself, see Are Zotero plugins safe?

What we check in the code

Code transparencyCan the code be read, and was the release file built from the public source?Sets the grade
Where your data goesWhich servers it contacts, and whether they're named services or the developer's own.Sets the grade
Powerful capabilitiesNative code, launching programs, running code built at runtime, reading files, storing passwords.Sets the grade
Works withWhich Zotero versions it installs on and runs in.Shown, not graded
What you'll needAPI keys, accounts or other software.Shown, not graded
MaintenanceHow recently it was released, and by how many people.Shown, not graded
LanguagesDocumentation and interface languages.Shown, not graded

Every finding links to the file and line it came from. The rules are published with the code, and they're the same for every plugin: when we change one, every card is recomputed.

The live test, for A+

Reading code shows what a plugin could do. The live test shows what it actually does.

  1. 1We install the exact release file in a fresh copy of Zotero 10, with a small sample library.
  2. 2We select items, open the reader and the plugin's settings, and click its menu items.
  3. 3We record every server it contacts and what it sends, including any of the sample library's text.
  4. 4If it loaded, and everything it did is something its card already describes, an A becomes A+.
Grades belong to versionsThe grade applies to the release named on the card. Each new release is checked again; if it adds a problem, the grade drops for that version and the change appears in the public log.
Think we got it wrong?Developers can respond to any card, and their response is published next to it with our reply. Anyone can report a problem we missed.Respond as the developerReport a problem

Changes to the checks

When the analysis or the rules change, every card is recomputed, and any grade that moves is listed in the grade log as "Our checks changed". This is what changed, newest first. The analyzer reads the code; the rules turn its findings into a grade.

  1. 13 Nov 2026 · analyzer 0.13.1 · rules 0.9.1-previewPackage installs: a message that describes a command (a log label such as "uv pip install (wheel)") no longer counts as one, and a folder or package named in a setting is the user's choice, not the plugin's supply chain. A wheel whose file name the plugin builds at runtime counts as one fixed version, options a plugin adds to a command after building it (args.push("--constraint", …)) are read, and exact packages held to a fully pinned constraints file the plugin ships count as a locked set (medium) instead of unpinned (high). aka.ms and download.visualstudio.microsoft.com are identified as Microsoft's download links. Of the 35 plugins whose installs we re-checked, one card changed: Zocrates (C to B, after its developer's response).
  2. 5 Nov 2026 · analyzer 0.13.1 · rules 0.9.1-previewzotero-plugin-toolkit 6.0.0: the scanner reads its debug link where 6.0.0 sets it up, counts it only when the plugin calls register(), and rates it high when the plugin turns the link's password and prompt off. No listed plugin uses 6.0.0 yet, so no card changed. Cards with the toolkit's install link say the toolkit removed it in 6.0.0. A card describes the file the plugin's own update address serves when that's newer than its latest GitHub release, and among several builds for the same Zotero the one the update address serves.
  3. 3 Nov 2026 · analyzer 0.13.1 · rules 0.9.1-previewA zotero:// link that switches the catalogue a plugin installs from is rated by what installing from that catalogue then does: medium while each install waits for a click or for a setting that's off by default, high when installs run by themselves. Before, it rated high like a link that installs an add-on without asking, which was stricter than the toolkit's install link (medium) even though the user still has to click.
  4. 3 Nov 2026 · analyzer 0.13.1 · rules 0.9.0-previewA text-direction control character inside a lookup table (a language dictionary, for example) no longer counts as hidden code: with only data after it, nothing is shown out of order. Keys in addresses a plugin builds at runtime are redacted in the evidence, like other keys.
  5. 16 Oct 2026 · analyzer 0.13.0 · rules 0.9.0-previewA program a plugin downloads and runs is rated by who can change it and whether anything checks it, and the card gives the reason. Until we've reviewed one, the finding stays high.
Earlier changes (15)
  1. 16 Oct 2026 · analyzer 0.13.0 · rules 0.8.0-previewA review also carries over for code outside any function and for shell scripts a plugin ships, as long as that code is unchanged.
  2. 16 Oct 2026 · analyzer 0.12.0 · rules 0.8.0-previewA review of a finding now applies to later releases whose code behind it is unchanged, and any finding can be reviewed, not only downloads. Plainer wording for AI assistants that run code (whether they ask you first) and for downloaded programs.
  3. 1 Oct 2026 · analyzer 0.11.0 · rules 0.7.1-previewMore addresses traced: ones stored on a class, and defaults behind a settings helper. A default server counts as where data goes, and the address is marked as one you can change.
  4. 1 Oct 2026 · analyzer 0.10.0 · rules 0.7.1-previewBundled libraries: the settings that decide where a library connects are read, so its own servers count only when the plugin turns them on, and the card says what a library loads for your content (an editor that shows videos, for example).
  5. 1 Oct 2026 · analyzer 0.9.0 · rules 0.7.0-previewCode counts as a known library's only where it matches that library's official release. A name or banner alone no longer exempts it; the rest is checked as the plugin's own.
  6. 1 Oct 2026 · analyzer 0.8.0 · rules 0.7.0-previewA plugin whose network code we can't trace is no longer graded B straight away. A quiet run in a test copy of Zotero, or a review of its network code, settles it; until then it shows as not graded yet ("still checking").
  7. 26 Sep 2026 · analyzer 0.8.0 · rules 0.6.1-previewTested in Zotero: when we've run a plugin's release file, the card says what the test did and found, and the hosts it sent data to are added to data sharing, marked "seen when we ran it", and count as requests we traced; pages it only loaded are listed but don't count as sharing, except on a usage-tracking service.
  8. 26 Sep 2026 · analyzer 0.7.0 · rules 0.5.0-previewNew findings: endpoints on Zotero's server and who can reach them, settings changed in Zotero and other programs, documents handed to other programs, browsers' and other apps' saved logins, where installs come from, data sent unencrypted. Packages rated by pinning, AI-written code by approval, the toolkit's zotero:// links only when set up.
  9. 26 Sep 2026 · analyzer 0.6.4 · rules 0.4.1-previewA second precision check (all ten labels right); keys saved in loops, download sources as requests, server sockets told apart from stream pumps.
  10. 26 Sep 2026 · analyzer 0.6.3 · rules 0.4.1-previewAddresses followed through request helpers and matched by scope; links through helpers; unused array entries and RDF namespaces aren't destinations.
  11. 26 Sep 2026 · analyzer 0.6.2 · rules 0.4.1-previewEvery confirmed review finding re-checked against the new cards. A public relay counts only when a request reaches it; more ways of loading code and downloading programs are followed; a Python program that opens Zotero's database counts; webhooks into the developer's own account; Zotero's server opened to the network; turning on Zotero's local API; about 35 more hosts classified.
  12. 26 Sep 2026 · analyzer 0.6.0 · rules 0.4.0-previewA precision check of ten random cards and a review of every "Serious concerns found" card. Commands a plugin only shows no longer count; downloads pinned in shipped scripts are pinned; update checks compare versions and read update.rdf, and a newer version raises the card only when it's worse; obfuscated code can't get a clean data-sharing result. Badges name the bundled library and add a detail line; the label lists the findings that set it; "contacts" only for traced requests.
  13. 26 Sep 2026 · analyzer 0.5.0 · rules 0.3.0-previewA full review of every card against its code. New checks: link handlers, encrypted code, download-and-run through any route, compiled programs and packages it runs, AI tools it hands data to, security settings, its own updater and servers, translators. Obfuscation confidence tiers. Update addresses fetched and followed. Web scripts in plugin pages lowered to low after a test in real Zotero showed Zotero blocks them.
  14. 26 Sep 2026 · analyzer 0.4.0 · rules 0.2.0-previewFixes from an independent review of the scoring: the build current Zotero installs, database writes through variables, bundled libraries attribute findings but never remove them, update addresses shown on their own line.
  15. 25 Sep 2026 · analyzer 0.3.2 · rules 0.1.0-previewFirst run over every listed plugin.

The full method, rule by rule: docs/methodology.md on GitHub.