Jev in Claude Code: what you can build

Yes, beside it, not inside it. TypeSafe's documentation says Jev is not a coding-agent language model: it returns a structured answer (a choice with probabilities, a score on a rubric you define, or a yes or no probability) and writes no text or code, so it cannot replace the model behind Claude Code.

What it can do is help in three ways: teach Claude Code to write programs that call Jev, act as a fast judge on the commands Claude Code wants to run, and sit inside agents you build as a cheap check between steps. The sections below take them in turn. The overview of every agent is on the main page.

• Jev in Codex: set it up and trust the hook
• Jev in OpenCode: skill, approvals and Zen

Claude Code installs TypeSafe's skill as a plugin. Run both commands, one after the other. TypeSafe's page warns against mixing install methods, because two methods leave duplicate copies.

claude plugin marketplace add typesafe-ai/skills

claude plugin install typesafe@typesafe-ai

To use it, type the command below, or name the skill in your prompt, for example "use the TypeSafe skill". Skills run with Claude Code's permissions, so read the skill file before you install it.

/typesafe:typesafe-ai

To update, run these two commands, then restart Claude Code or run /reload-plugins:

claude plugin marketplace update typesafe-ai

claude plugin update typesafe@typesafe-ai

It gives Claude Code the context to design Jev integrations. The typesafe-ai/skills repository, MIT-licensed and showing 2.8 thousand stars when we read it, contains one skill, typesafe-ai. It covers how to describe the state, how to write questions, how to compose a request, how to probe live and where to find the current documentation.

TypeSafe's page suggests three kinds of prompt. Ask Claude Code to find places where one judgment could replace complex parsing or fragile code. Export TYPESAFE_API_KEY and ask it to run cheap test queries and propose changes. Or point it at a TypeSafe cookbook and ask which pattern would make your code less fragile.

• How to write questions Jev answers well

OpenRouter's cookbook shows how. A script runs on every PermissionRequest for the Bash tool, which you register in .claude/settings.json with the command bun ./path/to/jev-permission-hook.ts and a timeout of 15. A fixed risk list in code goes first: recursive deletes, force pushes, sudo, publishing, deploy commands and credential paths such as .env or .ssh always go to the normal prompt and Jev is never called.

Everything else goes to Jev as two yes or no questions: is the command reversible and kept inside the project, and does it serve the task. The command is approved only if both scores are at least 0.9. Any error, timeout or unclear answer leaves the normal prompt in place. The cookbook measured about $0.0000168 for a request of roughly 400 input tokens, or about $0.017 per 1,000 checks.

Three Claude Code details from the cookbook matter. The hook receives Claude Code's one-line description of the command in place of a task, so the second question checks the command against that description. The hook does not run where no prompt can appear: in claude -p and in background subagents, a request with no hook decision is denied. And deny and ask rules in your settings still beat an allow from the hook.

Three uses that TypeSafe and its partners have published, each with the numbers they report. All figures are theirs and not independently checked:

• Pick the right skill before each turn. TypeSafe's skill-suggestion cookbook runs two Jev requests before every turn and adds the best skill to the prompt. On 488 test requests over 182 skills, with a Claude Haiku 4.5 agent, it reports wrong skill loads falling from 16.8% to 7.3% and needless loads from 9.8% to 4.0%. The run used Jev 1.12 and is dated 31 July 2026 in the cookbook.
• Decide whether an agent should take another step. Vercel's published example asks Jev "should the agent take another step?" as a yes or no question and continues when the probability is at least 0.8. It is a cheap stop-or-go check between steps.
• Let an agent propose better questions. TypeSafe's feature-discovery cookbook has an AI agent rewrite Jev questions over five rounds. On 2,000 wine reviews it reports an error of 1.772 points against 2.466 for word counts, using Claude Sonnet 5 as the proposer.

• Jev builds for coding agents

TypeSafe's page and the cookbook list these snags:

• The skill is not used. Invoke it explicitly with the command above, confirm the installer targeted Claude Code, then restart.
• Claude Code invents request fields. The plugin may be out of date. Update it and try again.
• Routing is off. Move the thresholds: too high causes missed cases and too low causes false alarms. Make each question narrower.
• The hook approves something you would decline. Add that command to the fixed risk list first. Jev sees only the request text, so the list, not Jev, is the real safety boundary.
• Scores move a little between identical requests. Start at 0.9 and lower toward 0.8 only if safe commands keep prompting.

• Where Jev struggles