# Jev Directory

> Everything Jev (typesafe-ai/jev) can do, with receipts: 50 runnable judge-model evals plus 1362 real community builds, each linked to its project and source post. Generated 2026-09-21.

## Start here: agents

- MCP server (preferred, always current): POST JSON-RPC 2.0 to https://jev.magicteams.ai/mcp — tools search_jev, get_jev_eval, get_jev_build, list_jev_categories, get_jev_pack, explain_jev, get_jev_integration_guide, list_jev_patterns, recommend_jev_use_cases, get_jev_eval_manifest; resources jev://evals, jev://guide, jev://playbook.
- [Capability pack (full content)](./capabilities.md): the whole playbook — how to call Jev, all 50 evals with runnable prompts, every build by category.
- [Setup prompt](./setup.txt): one paste that points an agent at the pack.
- [Structured data](./capabilities.json): evals, builds, model facts, and the eval manifest (rev 6d1f82b05fbea65e).

## Guides

- [What is Jev?](./what-is-jev.html) ([markdown](./what-is-jev.md)): judge-model explainer with the exact AI SDK call shape.
- [Jev, explained like you're 10](./jev-like-im-10.html) ([markdown](./jev-like-im-10.md)): the same idea with no jargon.
- [Decision patterns](./decision-patterns.html) ([markdown](./decision-patterns.md)): ten reusable judge-model patterns, each linked to runnable evals.

## Directory

- [Browse all builds](./index.html): search, categories, newest-first, copy-a-brief per card.
- [Sitemap](./sitemap.xml): every build page URL (individual builds are HTML only — prefer MCP search_jev / get_jev_build over crawling them).

## Key facts

- Model: typesafe-ai/jev via the Vercel AI Gateway (https://ai-gateway.vercel.sh/v1); auth AI_GATEWAY_API_KEY (vck_…) in the environment.
- Call: experimental_evaluate({ model, state, questions }) from the AI SDK; boolean / choice / score questions over one shared state.
- Price: $0.042 per 1M input tokens; output free. Eval manifest rev 6d1f82b05fbea65e: Exact match per question against the fixed state; an eval passes iff every question matches. Scores use the rubric scale (1 = first criterion … N = last); booleans and choices match literally.

## Contribute

- [Repo](https://github.com/everyai-com/jev-directory): source of this directory; [submit a build](https://github.com/everyai-com/jev-directory/issues/new?template=submit-use-case.yml).