How to use Jev in Claude Code, Codex and OpenCode

No. TypeSafe's own 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/no probability) and does not write text or code, so it cannot replace the model behind Claude Code, Codex, OpenCode, Cursor or Copilot.

What it can do is work alongside them. The documentation's recommended path for coding-agent users is to install TypeSafe's agent skill, which gives the agent the context it needs to write programs that call Jev.

• What is Jev?

There are three realistic uses, in order of how little setup they need:

• Help the agent write Jev code. Install the TypeSafe skill and the agent learns the API, the three question types and the usual patterns, so it can find places where one judgment call replaces fragile parsing and write the integration for you.
• Act as a fast judge while the agent works. A small hook sends each pending command or action to Jev and gets back a probability. OpenRouter publishes a worked example for approving routine shell commands.
• Run decisions inside the software the agent builds. Routing, scoring and checking statements are the uses TypeSafe's documentation names, and its patterns pages cover confidence routing and intent routing.

• Use cases

TypeSafe's agent-skill page gives one command per setup. In Claude Code, run claude plugin marketplace add typesafe-ai/skills and then claude plugin install typesafe@typesafe-ai. For Codex, OpenCode and other agents, run npx skills add typesafe-ai/skills --skill typesafe-ai and select your agent when asked. The install is project-local by default, and adding -g makes it global. You can also copy the skill's folder into your agent's skills directory by hand.

The page warns you to choose one installation method, to avoid duplicate copies. To update in Claude Code, run claude plugin marketplace update typesafe-ai and claude plugin update typesafe@typesafe-ai, then restart or run /reload-plugins. Elsewhere, npx skills update does it.

To use it, type /typesafe:typesafe-ai in Claude Code, or name the skill in your prompt in any agent, for example "use the TypeSafe skill". Skills run with your agent's permissions, so read the skill file before you install it.

The skill page suggests three kinds of prompt:

• Find opportunities. Ask the agent to find places where a judgment could replace complex parsing or fragile code.
• Run experiments. Export your API key as TYPESAFE_API_KEY (from the TypeSafe console) and ask the agent to run cheap test queries and propose changes.
• Apply a cookbook. Point the agent at a specific TypeSafe cookbook, or the cookbook index, to find patterns that make your code less fragile.

• How to write questions Jev answers well

OpenRouter's cookbook shows how. A small TypeScript hook runs before each shell command in OpenCode, Claude Code, Codex CLI or Cursor. A fixed risk list goes first and sends dangerous commands straight to the normal prompt: recursive deletes, force pushes, sudo, publishing, deploying and credential access. Everything else goes to Jev as two yes/no questions: is the command reversible and contained to the project, and does it serve the task. The command is approved automatically only when both scores are at least 0.9. Errors, timeouts and unclear answers fall back to the normal prompt.

The cookbook estimates about $0.0000168 per request, using Jev 1.13 through OpenRouter. It is also candid about the limits: Jev only sees what is in the request, so the fixed list is the real safety boundary, scores vary a little between runs, and approvals apply to one request and do not create saved rules. In the Codex setup it tested, the hook did not receive the API key from the environment, so it recommends a git-ignored .env file.

Yes, for anything that calls Jev. The skill's experiments use a key from the TypeSafe console, and TypeSafe currently asks for credit before it creates one. The permission-prompt cookbook uses an OpenRouter key. OpenCode's Zen service lists a free Jev 1.13 model for a limited time, which can be called from inside OpenCode, but it is OpenCode's offer, not TypeSafe's.

Installing the skill itself costs nothing. At Jev's list price of $0.042 per million input tokens with free output, the test queries an agent runs while you build are a small cost.

• Can you use Jev for free?
• Cost calculator

TypeSafe's page lists the common problems and fixes:

• The skill is not used. Invoke it explicitly, confirm the installer targeted your agent, then restart the agent.
• The agent invents request fields. The skill may be out of date. Update it and try again.
• Routing is off. Adjust the thresholds: too high causes missed cases and too low causes false alarms. Make the questions more specific.
• The code is hard to review. Keep the questions and the threshold values together in one file.

• Where Jev struggles

Our read: the useful picture is Jev as a cheap judge that your code calls, not a smarter coding model. Let your agent write the integration with the skill, keep one narrow yes/no or choice question per decision, and keep hard safety rules in ordinary code. Start by asking the agent where a judgment could replace a regex, because that is the lowest-risk first change.

• How to use Jev
• Jev vs ChatGPT, Claude and other LLMs