Use Jev with Codex, Claude Code and Coding Agents
Install TypeSafe’s agent skill, define focused implementation tasks, and keep keys and execution permissions separate.
On this page
What the TypeSafe skill doesInstall one methodGood tasks to delegateExample project promptKeep API keys privateDo not let the classifier execute actionsUpdating and troubleshootingWhat the TypeSafe skill does
TypeSafe provides a skill that teaches coding agents the API, primitives, patterns, and state-design conventions. Installing it adds implementation context; it does not by itself provide an API key or authorize actions in your project.
Install one method
For agents supported by the skills installer:
npx skills add typesafe-ai/skills --skill typesafe-ai
Choose your agent when prompted. The documented default is project-local installation; global installation is a separate choice. For Claude Code’s plugin route, TypeSafe documents:
claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-ai
Use one method to avoid duplicate versions. Review the skill source and its referenced files before installation. These commands are instructional examples; this website does not run them for you.
Good tasks to delegate
Ask an agent to find a bounded judgment in your application, propose a question schema, and write fixture tests before making live requests. Useful tasks include replacing brittle routing keywords, adding a review route, or evaluating passage relevance. Require the agent to preserve deterministic permission checks.
Example project prompt
Use the TypeSafe skill to add support-ticket triage.
Keep department, urgency, and refund detection as separate questions.
Create fixtures and a command that runs without an API key.
Do not execute refunds. Return a queue and review flags only.
Keep model version, questions, and thresholds in one reviewable module.
Before live evaluation, explain the request volume and expected cost.
Keep API keys private
Set TYPESAFE_API_KEY in the agent’s process environment or an ignored local secret file. Never paste it into a prompt, committed example, browser bundle, screenshot, or public issue. Review generated logs for accidental payload or header exposure.
Do not let the classifier execute actions
A label such as allow is a recommendation. The host still enforces permissions, exact arguments, and confirmations. Test a malicious proposal that asks the model to ignore its policy; the deterministic executor must remain protected.
Updating and troubleshooting
A stale skill can produce incorrect field names. Compare generated calls with the current API guide and official reference. Use instructions and criteria, not guessed fields. TypeSafe documents npx skills update for installer-managed copies and separate plugin updates for Claude Code.