Jev in OpenCode: skill, approvals and Zen

Yes, beside it. TypeSafe's documentation names OpenCode among the coding agents Jev cannot replace, because Jev returns typed decisions and writes no code. It can help OpenCode write programs that call Jev, approve routine shell commands, and be called directly through OpenCode's own Zen service.

• Jev in Claude Code: what you can build
• Jev in Codex: set it up and trust the hook

Run TypeSafe's installer and choose OpenCode when it asks. The install is project-local by default; add -g for a global install. Use one install method only.

npx skills add typesafe-ai/skills --skill typesafe-ai

Then say "use the TypeSafe skill" in your prompt. To update, run this:

npx skills update

OpenRouter's cookbook covers OpenCode 2.0.12 or later; it does not cover version 1. First, in opencode.json, set "shell": "ask" under permission, so that shell commands prompt and the plugin has something to judge. Then save the plugin file under .opencode/plugins/ in your project. It loads with no install step, and opencode plugin list should show jev-permissions as a local plugin.

The flow is the one the other agents use. A fixed risk list in code sends dangerous commands, such as recursive deletes, force pushes, publishing and credential paths, straight to the normal prompt. Everything else goes to Jev as two yes or no questions, and the command is approved only if every score is at least 0.9. Any error or timeout falls back to the prompt. You need OPENROUTER_API_KEY in the environment OpenCode runs in.

OpenCode details from the cookbook: deny rules in opencode.json apply before the plugin runs, so put commands that must never run there. Approvals apply to one request and write no saved rule. The plugin reads the commands from event.resources, not event.metadata. To test, ask for a change that runs tests, which should run without a prompt, then ask it to push and confirm the push still prompts.

Yes, on Jev 1.13 only. OpenCode's Zen page lists jev-1.13, at $0.042 per million input tokens with free output, and jev-1.13-free, free for a limited time with no end date given. The page documents a direct call to https://opencode.ai/inference/systemone/v1/systemone with a Zen key as a Bearer token, in the same request format as TypeSafe's API. It does not say how to pick Jev inside OpenCode's terminal app.

The free model is OpenCode's offer, not TypeSafe's, so it can end without notice, and the setup steps ask you to add billing details before copying a key. Avoid sending private data to it.

• Can you use Jev for free?

Ideas that TypeSafe has published as cookbooks, with the figures it reports. They are TypeSafe's own and not independently checked:

• Pick the right skill before each turn. TypeSafe's skill-suggestion cookbook, run on 182 skills and 488 requests, reports wrong skill loads falling from 16.8% to 7.3%. It used Jev 1.12 and a Claude Haiku 4.5 agent.
• Choose a cheaper model when the answer is easy. OpenRouter's Jev Router (typesafe/jev-router) uses Jev to pick the model and reasoning effort for each request, with three cost tiers: low, medium and high. OpenRouter gives no price for the router itself.
• Check whether a cheap answer holds up. OpenRouter's cascade cookbook has a cheap model draft an answer, Jev check that the sources support it, and a stronger model run only when the check fails. On 50 questions it reports no wrong answers at $0.012, against two wrong at $0.175 for the stronger model on every question.

• Jev on OpenRouter

The snags the documentation lists:

• The plugin never fires. Check that shell is set to ask, the file is under .opencode/plugins/ and opencode plugin list shows it. Version 1 of OpenCode is not covered.
• A deny rule blocks a command you expected Jev to approve. Deny rules run first and always win.
• The skill is not used. Name it in your prompt, confirm the installer targeted OpenCode, then restart.
• Zen's free model disappears. OpenCode says it is available "for a limited time". Plan for it to end.