Search for a Jev plugin and you will find a dozen repositories with thousands of stars between them. One of them is official. The rest are community projects that solve real problems and carry no support commitment from TypeSafe, and mixing the two categories is how teams end up with a production dependency nobody is maintaining.
The distinction is not about quality. Some community projects are better tested than internal tooling you already run. It is about knowing what you are installing, who answers when it breaks, and what leaves your machine.
TypeSafe maintains one agent skill, MIT licensed, and its own README lists a single skill in the table. That skill teaches a coding agent how to design with Jev: find current documentation and cookbooks, and compose typed judgments in code. It is installed through the Claude Code plugin marketplace or through `npx skills add` for other agents, both documented in the installation guide.
That is the whole official surface for agent plugins. The Python and JavaScript SDKs are official libraries rather than plugins, and the System One adapter is official too. If a repository is not under the `typesafe-ai` organization and is not one of those, treat it as community work regardless of how it is described.
The projects that matter cluster around one problem: context handling in a coding agent, where a long session gets compacted and detail is lost. TypeSafe's Jev is a good fit for that decision because "which tool results are still needed" is a bounded judgment with a yes or no answer per item. That is why compaction is where most of the ecosystem went.
Two observations from that table. Every project is MIT licensed, which means you can read and fork all of them. And several were pushed within days of each other, which tells you this is an area under active churn rather than settled infrastructure. The compaction comparison covers the design trade-offs between these approaches in more detail.
The largest community project describes itself as a Claude Code function-hook plugin, and its README shows what that means in practice. It needs a Claude Code version at or above 2.1.274, it needs function hooks enabled through `CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1`, and it reads `TYPESAFE_API_KEY` from the environment like the official path does. Run from a checkout, and it is launched with a plugin directory flag.
Three consequences follow. The hook mechanism is early access, so a Claude Code update can change behaviour underneath the plugin. The key handling is your responsibility, and the README says so directly. And the plugin is positioned to fail with a fitted throw that lets the caller decide the fallback, which is a deliberate design choice worth understanding before you enable it on a repository you care about.
Do these five things before a Jev plugin touches a repository that matters.
That is the same discipline the tool-call gate and bounded permissions posts describe for agents generally. A plugin that rewrites context is an agent component with write access to your agent's memory.
Is there an official Jev plugin for OpenCode?
TypeSafe's official skill works across skill-compatible agents through the `skills.sh` install, and the OpenCode walkthrough covers the host-specific configuration. The OpenCode compaction project in the table is community work.
Do community Jev plugins need their own API key?
They use the same `TYPESAFE_API_KEY` convention as the official path, which is convenient and also means one credential is now spread across more tools. Keep it out of committed configuration files.
Can I use several Jev plugins at once?
Technically yes, and two compaction plugins enabled together is a bad idea because they both own the same decision. The routing explainer is the useful mental model: separate the decisions, then give each one exactly one owner.
Should I install a community plugin instead of the official skill?
They are not alternatives. The official skill changes how your agent writes and designs. A community compaction plugin changes how your agent manages a long session. Most teams that get value from the ecosystem run the official skill and one plugin, chosen deliberately.
The hard part is not installing a plugin. It is deciding which decisions in your agent deserve a typed judgment, which deserve code, and which deserve a person, and then keeping that boundary intact as the plugin count grows. That is an engineering decision with a six-month consequence.
Maxpertise embeds an AI engineer inside your existing team to make those decisions with you and review them before they land, on our services page, with a written intake as the way in. If you are weighing the cost of that against doing it in-house, start with what an AI engineer costs.
[1] TypeSafe, typesafe-ai/skills, accessed October 12, 2026. Read directly: one skill listed in the repository's own table, MIT licence, and the statement that it is for "building with TypeSafe". This is the only plugin surface published by the TypeSafe organization as of that date.
[2] TypeSafe, Agent skill guide, accessed October 12, 2026. Installation, invocation and update paths for the official skill across Claude Code and other skill-compatible agents.
[3] tamaratran, fast-jev-compaction, repository metadata and README read through the GitHub API on October 12, 2026: 7,594 stars, MIT, last push 2026-09-18. The README describes the repository root as "a Claude Code function-hook plugin: `hooks/fast-jev.ts`", states that function hooks are an early-access Claude Code feature requiring 2.1.274 or above, documents `CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1` and `TYPESAFE_API_KEY`, gives `https://api.typesafe.ai/v1/systemone` as the default `baseUrl`, ships `tests`, `demo` and `examples` directories, and describes a failure mode where the hook throws and "the caller (or the Claude Code hook) decides what to fall back to".
[4] cth9191, jev-compaction-plus, read through the GitHub API on October 12, 2026: 26 stars, MIT, last push 2026-10-02, described as a fork of fast-jev-compaction that keeps retained context word for word and "moves the rest to a drawer file". Its speed comparison is the project's own claim, not a measurement made here.
[5] overbit, fast-jev-compaction-opencode, read through the GitHub API on October 12, 2026: 13 stars, MIT, last push 2026-10-03, described as Jev-guided context compaction for OpenCode that prunes stale tool history rather than rewriting conversations.
[6] jonathanavis96, jev-kit, read through the GitHub API on October 12, 2026: 63 stars, MIT, last push 2026-10-02. Described as a tool-call guard, tier guard, file search, browser agent, review, belay, compaction and installers.
[7] kerpopule, hermes-jev-skills, read through the GitHub API on October 12, 2026: 1,092 stars, MIT, last push 2026-10-09. Described as Jev-powered model routing, memory, compaction, skill selection, and computer and browser use for Hermes agents, also Claude Code and Codex.
Maxpertise is an AI-native engineering company. We embed native AI engineers inside your team, live in about 10 days. Please enable JavaScript to view the site, or email contact@maxpertise.net.