claujev

English · Português · Website · Patch notes

ClauJev plugs Jev, TypeSafe AI's fast judgment model, into Claude Code.
While Claude works, Jev answers short, typed questions such as "is this command destructive?" or
"was the request fully done?", and ClauJev turns the answers into action:

• 🧭 Triage: tells Claude whether a request is a question, a small change, a big one, or too vague.
• 🎚️ Automatic effort: low for a quick question, xhigh for a tangled bug, chosen per request.
• 🛑 Command guard: asks for your confirmation before destructive or credential-touching commands.
• 🔍 Rule review: checks every edit against rules you write in plain language.
• ✅ Final check: stops "done!" when the work looks incomplete or your tests fail.
• 🪄 Subagent routing and 🕵️ false-claim detector for subagents.

About 400 ms per check. Six days of daily use cost the author US$ 0.07.

It can also save Claude tokens. Jev costs a fraction of a cent, and it steers Claude away from the
expensive habits: low effort (less thinking) on simple questions, "go straight to the file" instead
of exploring the whole repository for a one-line change, no edits when you only asked a question, and
simple subagent tasks on a smaller, cheaper model. The final check also catches unfinished work before
you have to ask again.

[!NOTE]
Community project, not affiliated with TypeSafe AI or Anthropic. You need a
TypeSafe AI API key.

Install

In Claude Code:

/plugin install claujev --marketplace Civalski/claujev

Enter your TypeSafe AI key (the only thing the install asks) and restart Claude Code. On the first
session Claude offers a short guided setup in your language; run /claujev:configure to change it later.

Optional, after the step above: /plugin install claujev-mod --marketplace Civalski/claujev for the
status line and automatic effort. (Installing the mod first also installs ClauJev, but without asking
for the key: then run /plugin configure claujev@claujev.)

Commands

You can also just ask Claude: "check that ClauJev works", "set up ClauJev for this project",
"show me the ClauJev report".

How it works

ClauJev listens to six Claude Code hooks. For each, it builds a small state (the request, the
command, the diff), asks Jev the right questions and hands Claude an action, or nothing.

flowchart TB
    U(["👤 You"]) -->|request| CC["🤖 Claude Code"]
    CC -->|"hook event"| CJ["⚙️ claujev.mjs<br/>one file · zero deps"]
    CJ <-->|"questions ⇄ probabilities<br/>~400 ms · secrets redacted"| JEV[("🧠 Jev · TypeSafe AI")]
    CJ --> R{"what goes back<br/>to Claude Code?"}
    R -->|"prompt"| A1["💬 hint + effort"]
    R -->|"bash"| A2["✋ ask for your confirmation"]
    R -->|"edit · stop · subagent"| A3["🔁 review before continuing"]
    R -->|"agent"| A4["🪄 smaller model"]
    R -->|"all low"| A5["✅ carry on"]

• Never breaks the session: any error is logged and Claude carries on.
• Observe before acting: each answer falls into a low / uncertain / high band; only high acts, and only in enforce.
• A signal, not a verdict: ClauJev never deletes or reverts anything. A false alarm costs one sentence from Claude.
• Sees the real work: the final check reads the turn's diff and the commands Claude ran (commits, pushes, releases).

Automatic effort

flowchart LR
    P["request"] --> J{"Jev triage"}
    J -->|"question / trivial"| L["low"]
    J -->|"small change"| M["medium"]
    J -->|"medium change"| H["high"]
    J -->|"large change<br/>or ambiguous bug"| X["xhigh"]

With claujev-mod, observe (default) shows the suggestion in the status line and on applies it,
within the limits you set. max is never picked automatically, subagents keep their own effort, and a
request that asks for effort itself ("ultrathink") is never overridden.

Privacy

Sent to api.typesafe.ai: the request (and the previous one, as context), commands, edited snippets,
the turn's diff and commands, and Claude's final reply of the turn (up to 3,000 characters), after
removing keys and tokens. Kept locally in ~/.claude/claujev/: logs with answers and a content
hash, never the content. The API key lives in your system's secure storage.

Learn more

• Configuration guide: project rules, areas, verify, your own questions, every option, decision bands, the report, manual install
• What ClauJev asks Jev: the exact questions per hook
• Changelog

Does it slow Claude down? About 400 ms per check; read-only commands skip Jev entirely.

What if Jev is wrong? Actions are reversible by design, thresholds are high on purpose, and the
uncertain band never acts. Start in observe and read the report.

Does Jev see the whole conversation? No, only the state built for each question.

And the official typesafe@typesafe-ai plugin? That one teaches Claude to build with Jev.
ClauJev uses Jev to supervise Claude. They work side by side.

Contributing

Issues and PRs are welcome. npm test runs the core tests, claude plugin test plugins/claujev-mod
the mod's, npm run validate the manifests.

License

Created by Alisson Civalski. MIT: free to use, modify and
redistribute, including commercially, as long as the copyright notice is kept.