How to use Jev with Claude Code
You cannot make Jev the model that powers Claude Code; TypeSafe says plainly that Jev is not a drop-in replacement for the LLM behind a coding agent (Jev with coding agents). What you can do is install TypeSafe's official agent skill so Claude Code writes correct Jev code, or add community hooks and plugins that call Jev for small, typed decisions inside a Claude Code session.
Jev does not replace Claude Code's model
Coding agents like Claude Code rely on an LLM that streams text, calls tools and edits files. Jev does none of that. It takes a state and typed questions and returns a choice, a score or a yes/no probability. TypeSafe notes there is no model: "jev-latest" setting that turns a coding agent into a Jev-powered agent (Jev with coding agents).
That leaves two real options: use Claude Code to build software that calls Jev, or put Jev beside Claude Code's model to make narrow decisions for it. Jev vs an LLM explains the difference in more detail.
The official route: the TypeSafe agent skill
TypeSafe publishes an agent skill that gives Claude Code context on the Jev API, the three question types and TypeSafe's architectural patterns. In Claude Code it installs as a plugin with two commands (Agent skill):
claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-aiYou can then call it directly with /typesafe:typesafe-ai, or just say "use the TypeSafe skill" in a prompt. The skill lives in an MIT-licensed repo (typesafe-ai/skills).
The docs add some practical advice (Agent skill):
- Keep questions and threshold constants in one file, so people can review them easily.
- Expect to edit questions together with the agent, because "agents aren't great at writing questions."
- If the agent invents request or response fields, the skill may be stale. Update it and retry.
This skill helps Claude Code write Jev integrations. It does not add hooks or change how Claude Code itself decides anything. For a first integration, see how to use Jev.
Claude Code hooks with Jev: a community pattern
TypeSafe's docs do not document Claude Code hooks, so every Jev hook listed below is a community project. They share a common shape. A hook script collects a small state, such as the pending tool call or the transcript so far, asks Jev a few typed questions, and lets code decide what happens next. The projects below attach this to hook points such as PreToolUse (before a tool runs) and Stop (before Claude finishes a turn).
That shape matches TypeSafe's own design advice: keep control flow, deterministic rules and side effects in code, and give the model only narrow, typed questions (How to build with TypeSafe). The confidence docs suggest three bands: act automatically when confidence is high, proceed with caution when it is medium, and do not act when it is low. Thresholds should scale with the risk of the action (Confidence). In a hook, those bands map naturally to allow, ask and deny.
What people use Jev for inside Claude Code
- Guarding tool calls: scoring a command or edit for risk before it runs. See safety and guardrails.
- Trimming context: deciding which tool output, rules or history deserve a place in the context window. See context and memory.
- Picking skills and models: ranking installed skills or routing a turn to a cheaper model. See routing and model choice.
- Checking "done": testing whether the agent's claim of success matches what the session shows. See coding and code review.
TypeSafe has measured one of these jobs itself. In its skill suggestion cookbook, two Jev requests per turn ranked and re-checked 182 skills for the Hermes agent harness. Wrong skill loads by claude-haiku-4-5-20251001 fell from 16.8% to 7.3% (Skill suggestion). That test used Hermes, not Claude Code, but the problem is the same.
Cost, speed and limits in a hook
Hooks can fire on every tool call, so cost and rate limits matter. Jev 1.13 charges $0.042 per million input tokens, output is free, and the limit is 1,200 requests per minute. Over-limit calls get a 429, and the official SDKs retry with backoff by default (Models). Several of the projects below are designed to fail open, meaning they let the action continue if the Jev call fails, so an outage does not block your session. More detail is on Jev pricing.
Projects doing this
Agents & browsersTypeSafe agent skillsTypeSafe's official skill, which gives Claude Code, Codex and compatible agents context on Jev's primitives and patterns.GitHub · ★ 2.1k · typesafe-ai
Coding & code reviewbuilding-with-jev-skillA community skill for writing, fixing and improving programs that call Jev.GitHub · ★ 131 · dbreunig
Context & memoryFast Jev CompactionA Claude Code plugin that replaces the compaction summary with Jev decisions, scoring every tool call and result in one request and keeping what stays verbatim.GitHub · ★ 6.8k · tamaratran
Routing & model choicejev-router (gargpratyush)A local proxy that asks Jev to route each Claude Code or Codex turn to the cheapest model that can handle it, with a deterministic fallback.GitHub · ★ 377 · gargpratyush
Context & memoryjev-prunerA Claude Code plugin that trims long Bash output with Jev before the main model reads it.GitHub · ★ 144 · tamaratran
Context & memoryWinnowA context sieve for Claude Code where Jev judges each large tool result before it enters context.GitHub · ★ 86 · GhalebDweikat
Coding & code reviewCannyA warden for Claude Code and Codex CLI that uses deterministic hooks to stop an agent claiming work is done without evidence, with Jev as an advisor.GitHub · ★ 76 · qkal
Safety & guardrailsjev-guardA security hook that has Jev risk-score every tool call with session context (deny, ask or allow) and flag prompt injection in results.GitHub · ★ 32 · leepokai
Safety & guardrailsjev-guard (muratcakmak)A Claude Code plugin with PreToolUse-style hooks that combine regex rules and batched Jev noul scores to deny rule-breaking edits and unasked-for deploys.GitHub · ★ 7 · muratcakmak
Coding & code reviewThe Jev-enatorThree Claude Code hooks that each ask Jev a question: a PreToolUse danger gate, a PostToolUse failure notice and a Stop completion check.GitHub · jakenbear
Safety & guardrailsclear-headA Claude Code Stop hook that checks factual claims in the answer against what the assistant actually read in the session, with Jev as the judge.GitHub · ★ 2 · VladyslavHontar
Coding & code reviewjev-skill-gateA tool that scores installed Claude Code skills for relevance with Jev and writes skillOverrides so only useful ones reach context.GitHub · ★ 5 · ShivamPansuriyaQuestions
Can I set Jev as the model in Claude Code?
No. TypeSafe states that Jev is not a drop-in replacement for the LLM behind Claude Code or similar tools, and there is no model setting that makes a coding agent run on Jev.
How do I install the TypeSafe skill in Claude Code?
Run "claude plugin marketplace add typesafe-ai/skills" and then "claude plugin install typesafe@typesafe-ai". You can invoke it with /typesafe:typesafe-ai.
Does TypeSafe provide official Claude Code hooks for Jev?
No. The official offering is the agent skill. Hooks that call Jev, such as PreToolUse guards and Stop checks, are community projects.
What should a Jev hook do when Jev is unsure?
TypeSafe suggests acting on high confidence, proceeding carefully on medium confidence and not acting on low confidence, with higher thresholds for riskier actions. In a hook that usually means allow, ask or deny.
Will calling Jev on every tool call slow Claude Code down?
Each call adds a network round trip, and Jev 1.13 is limited to 1,200 requests per minute. Asking several questions in one request helps, because they are evaluated in parallel.
More guides: What is Jev?Jev pricing and API costJev vs LLMs: when to use whichHow to use Jev: a quickstartJev and MCP: using Jev as a Model Context Protocol toolJev in an agentic harness: where it fits in an agent loopJev for orchestration: routing requests, models and skillsJev as a judge: evaluating LLM and agent outputsJev for SEO and GEOJev for ads and ad reviewJev for marketing and lead scoringIs there an open source Jev?Can you fine-tune Jev?Designing questions for JevWhat is RLCD?Jev statistics