jev-recipes weekly downloads statistics

jev-recipes

Typed AI decisions for agents, RAG, and support flows: route, rerank, gate, grade, compare, and label text with one function call. Plug into Vercel AI SDK or LangChain, or call any recipe directly.

Recipe guides at jev-recipes.com | Quickstart | Live music app | Coding assistants | Recipe catalog | API vs. SDK vs. recipes | Agent frameworks | Contributing

248 focused recipes for JavaScript and TypeScript. Route messages, check evidence, and label model responses with a function call.

Documentation versions

Run npm ls jev-recipes in your project to check the version you installed, then choose its v tag from all release tags to browse matching documentation.

The hosted catalog and links to main can include unreleased changes. See release notes for compatibility and migration changes.

Help and contributions

Ask a usage question, report a software bug, share an unexpected model decision, or request a recipe.

For a reproducible software problem, start with the standalone reproduction example. Share sanitized inputs and logs, never credentials or private data. To help improve the project, see your first contribution.

How it works

Each recipe accepts your data, calls Jev through TypeSafe's System One API or an injected compatible client, and returns a structured decision. Use it in a Node.js backend, a script, or a research evaluation. Your application decides what happens next.

Start with one decision

| Your task | Recipe guide | Try without a key | Build the workflow |

| Route work to the right team | route | npx jev-recipes demo route | Route work | | Select useful evidence | rerank | npx jev-recipes demo rerank | Select evidence | | Review an agent's proposed action | tool-call-gate | npx jev-recipes demo tool-call-gate | Review an action |

These demos use saved responses. The runnable workflow examples show ready decisions, uncertainty, and provider failures. Use TypeSafe directly or Vercel Gateway for live calls.

For example, give route a support message and descriptions of your teams. It returns a team such as billing, or a review outcome when the choice is uncertain.

Evaluate your own cases with the installed evaluator, retain model responses, and replay confidence policies offline. The agent workflow and customer queue show how decisions fit into application code. Direct and Gateway integrations use the same recipe interface.

Watch Jev play checkers against Jev using the checkers-move recipe

Build a complete support flow

The support conversation starter checks for missing details, asks a configured question, and resumes with the customer's answer. It proposes a queue or requests human review when the conversation remains unresolved.

Open http://localhost:3000 and choose Ask and continue. This folder is portable and needs no repository build. Edit config.mjs to supply your queues, required information, questions, and review policy. Run npm run demo for the command-line version. Saved scenarios work without a key; the starter guide explains live calls and application integration.

The existing rules/Jev/fallback comparison covers single-turn routing, not this clarification loop. Its evidence remains experimental.

Build the searchable static catalog with npm run site:build, then preview it with npm run site:preview. It includes fixture exploration, related-recipe comparisons, and saved evaluation evidence. Current measurements and missing, unknown-origin, or older evidence are labeled explicitly. Use npm run eval:audit after building to check retained evidence offline; see evaluation and archive format.

Hear a real application: Jevthoven

Jevthoven is a music app built by this package's maintainer using jev-recipes. Its interface lets listeners choose a style, key, and tempo, and displays Jev's musical decisions alongside a piano roll. Playback requires sign-in.

Open the music app · Read the integration example

This is a maintainer-built application, not an independent customer adoption claim or a recipe accuracy benchmark.

Use a recipe

Requires Node.js 22.9 or newer, ES modules, and a TypeSafe API key. Live calls send the supplied input to TypeSafe and use API quota.

1. Install in your project folder.

2. Add the code below to your app, or save it as route-message.mjs in the same folder to try a standalone Node example.

The .mjs extension is optional. We suggest it for this example because Node recognizes it as an ES module without changing your project settings. Existing apps can use .js with "type": "module" in package.json, or .ts with their usual TypeScript setup. See Node's module formats.

3. Set your key and run. Replace your-api-key with your TypeSafe key. For the standalone example:

If you saved it as route-message.js in an ES module project, run node route-message.js. For an existing app or TypeScript project, use its normal start command.

Windows PowerShell

Example output:

A live result can differ. When the choice is uncertain, the script prints the review message instead.

Make it yours: replace request with your incoming message and routes with your team's names and descriptions. Replace console.log with your queue or handler logic. Keep the key and live calls on the server.

Try the saved example without an API key

This runs an offline fixture and prints JSON with result.status: "ready" and result.route: "billing". No model is called. Use it to inspect the interface before setting up a key.

Why recipes?

The TypeSafe SDK handles API calls and typed answers. Recipes add the instructions, input and response validation, confidence policy, and result handling for a specific decision.

| Start with | You provide |

| Raw API | HTTP handling, question, choices, validation, and review logic | | TypeSafe SDK | Question, choices, task-specific validation, and review logic | | jev-recipes | route({ request, routes }) and the application code that uses its result |…

Description from the npm listing.