User repository on GitHub
Jev (TypeSafe) model router on the Vercel AI Gateway
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Agents & MCP
Jev Model Router — az9713
Web chat where Jev picks the LLM for each message: one Choice question over four priced tiers, then the chosen model replies through the same Vercel AI Gateway key.
- Stars
- 2
- Language
- JavaScript
- Published by
- Community
- Author
- az9713
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is Jev Model Router — satviksinha.
About this project
A small web chat where Jev, TypeSafe's decision model, picks which LLM should answer each message. The chosen model then replies. Both calls go through the Vercel AI Gateway with one key. Inspired by Riley Brown's video Jev: the model that can't write, where he builds the same router in one prompt. This repo is that idea, built and documented step by step with Claude Code.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- JavaScript
Highlights
- One Choice question to typesafe-ai/jev with four options — nano, fast, balanced, frontier — each with a one-line rubric edited in the TIERS block.
- The response carries Jev's probability for every tier, its confidence, and the timings of both the routing call and the reply.
- One Jev decision is roughly 340 input tokens, about $0.000014 at $0.042 per million, on top of whatever the chosen tier costs to answer.
- npm test runs offline checks and the held-out routing cases for free; npm run eval:live re-runs them against Jev only, never a reply model.
Watch out
The tier table runs from $0.05 to $10 per million input tokens and $0.40 to $50 per million output, so a single mis-route to frontier costs far more than the routing decision ever will — the decision itself is roughly $0.000014. There is also a free-tier trap: the gateway's free plan limits Jev to a few calls and refuses most reply models, which makes the paid tier a prerequisite. The README publishes no routing accuracy, and the held-out cases run offline by npm test ship without results, so treat the rubric as unproven. Tiers and the question live in two blocks at the top of server.mjs, which is the only place routing behaviour is configured. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
The README does not claim it is good. It describes the project as inspired by a video in which the same router is built in one prompt, and this repo as that idea built and documented step by step. It does ship a harness: npm test runs offline checks plus the held-out routing cases with no credits, and npm run eval:live evaluates the same cases against Jev and writes one JSON object per case without calling any reply model. No accuracy figures are published, so the question of Jev-versus-heuristic stays open.
The routing call is noise: about 340 input tokens at $0.042 per million, roughly $0.000014 a decision. The reply is what costs — nano at $0.05/$0.40 per million, fast at $0.50/$3.00, balanced at $2/$10 and frontier at $10/$50 for input/output. So a wrong frontier pick on a long answer can cost hundreds of times more than the decision that made it, which is why router accuracy matters more than router latency here.
Not the chat itself. It needs Node 20.6 or later and a gateway key on the paid tier; the free tier limits Jev to a few calls and refuses most reply models. Several paths do work offline: node server.mjs --check is a free offline server check, npm test runs the offline checks and the held-out routing cases without credits, and node server.mjs --offline with python tests/browser.py exercises the browser flow against a stub. probe.mjs sends one raw Jev call and prints the full response body.
No threshold or reroute is documented. Jev returns a probability for every tier plus a confidence, and the server hands the reply, the probabilities and both timings back to the page, so a low-confidence pick is still executed by the model it named. The candidate set is a fixed four-item list, and changing it — or the question itself, including each option's one-line rubric — means editing the TIERS and TIER_QUESTION blocks at the top of server.mjs.
It needs Node 20.6 or later and a Vercel AI Gateway key on the paid tier: npm install, echo AI_GATEWAY_API_KEY=your_key > .env, node server.mjs --check for a free offline check, then node --env-file=.env server.mjs serving http://localhost:3000.
nano (openai/gpt-5-nano), fast (google/gemini-3-flash), balanced (anthropic/claude-sonnet-5), and frontier (anthropic/claude-fable-5.1). Jev sends one Choice question with those four options, each with a one-line rubric.
One Jev decision is about 340 input tokens at $0.042 per million, about $0.000014; the reply then costs whatever the chosen tier costs. npm test and the held-out routing cases run offline without spending credits.
More Jev resources like this
9GitHub stars
Local bridge where TypeSafe Jev picks the model and reasoning effort for Codex Desktop and CLI sessions, reusing the existing Codex login.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
ansidium/jev-codex-bridgeJavaScript
ansidium/jev-codex-bridgeJavaScript4GitHub stars
OpenClaw plugin adding a jev_route tool that classifies task type, complexity and risk with TypeSafe Jev, then maps the result to a route through testable JavaScript policy.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
trietphan/jev-clawJavaScript
trietphan/jev-clawJavaScript2GitHub stars
An MCP server packaged as both a Codex and a Claude Code plugin, exposing eight read-only tools that turn Jev probabilities into proceed/review recommendations for routing, ranking, extraction and claim checks.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
E-FL/typesafe-as-a-judgeJavaScript
E-FL/typesafe-as-a-judgeJavaScript2GitHub stars
Hooks that put Jev in two spots of a coding agent's loop, picking which skill a turn needs and which model tier a subagent gets, with deterministic rule gates in front and fail-open behaviour everywhere else.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
onlyjq04/jev-agent-hooksJavaScript
onlyjq04/jev-agent-hooksJavaScript7GitHub stars
A local MCP server that gives Codex a few cited excerpts instead of a whole noisy search result, with a deterministic default selector that needs no Jev account or key.
Jev Agents & MCP#community#jevbest#catalog-import
Community
suenot/codex-jev-routerJavaScript
suenot/codex-jev-routerJavaScript2GitHub stars
A Node CLI that reads a folder of Word, Excel, PowerPoint, PDF, email, CSV, HTML, code and text files, classifies each with Jev on OpenRouter, and writes a catalog, an INDEX.md and an offline report. Originals are never moved.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
nexibeo/jev-organizeJavaScript
nexibeo/jev-organizeJavaScriptFrom the Jev community
Posts from builders shipping with Jev right now.