Does it compute expected cost from model prices?

No. Jev returns a tier distribution such as mostly strongest, then the router applies confidence gates. PowerShell-style revenue math is irrelevant here; routing reads probabilities directly without price scaling.

How does two-model routing differ from three-model routing?

Two models use one threshold on strongest probability. Three start in the middle: escalate on strongest confidence, drop on cheapest confidence, otherwise stay middle, with escalate checked first.

Which way do I turn the knobs to save money?

Lower upgrade plus higher downgrade spends more freely; higher upgrade plus lower downgrade stays cheap. Moving both the same way is inconsistent rather than more extreme, and defaults lean to quality.

What happens when the Jev API itself fails?

It uses the configured fallbackTier model and marks isFallback true. Result still carries tier, model name, and the returned probability map for inspection.

How do I install and configure it?

npm install jev-model-router, get a TypeSafe key from the console and set TYPESAFE_API_KEY, then construct a Router with two or three models ordered weakest to strongest, each with a name and a capability description.

How many models can I route between?

Exactly two or three, ordered weakest to strongest. The count is read from models.length rather than configured, so passing threshold with three models or the cascade knobs with two throws at construction.

What do the tier descriptions actually do?

Each model's description is one level of the capability rubric sent to Jev, so the descriptions are what the classifier reads to place a query. The model names themselves are only labels on the tiers.

CommunityJev Agents & MCP

Jev Router — rajdhakad9826

Node router that sends each query to the cheapest capable LLM tier by asking Jev to classify demand against your tier descriptions, with no LLM call spent on routing.

See highlights

Open github.com

Stars: 10
Language: TypeScript
License: MIT
Published by: Community
Author: rajdhakad9826
Last verified: Oct 2026

rajdhakad9826/jev-router

User repository on GitHub

View on GitHub

LLM router that picks the cheapest model capable of handling a query, using TypeSafe's Jev for fast classification instead of an LLM call.

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

This is one of several Jev listings for the same project. The canonical entry is Jev Router — gargpratyush.

About this project

jev-model-router routes each query to the cheapest LLM tier that can actually handle it, using TypeSafe's Jev to classify how demanding a query is — fast and cheap, without spending an LLM call on the routing decision itself. jev-model-router needs a TypeSafe API key to call Jev. Get one from the TypeSafe console, then set it as an environment variable before constructing the router:

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Routing

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Tier descriptions form the capability rubric Jev scores; ordering weakest to strongest selects binary or cascade mode.
Confidence gates replace cost formulas after uneven price gaps left confident needs unselected in prior versions.
Strict validation rejects threshold fields that do not match the configured two- or three-model count.
Results expose model, tier index, full probability map, and fallback flag for logging and tuning.

Watch out

Models must stay ordered weakest to strongest with exactly two or three entries, mismatched strategy fields throw at construction, and escalate wins when both boundaries trigger. 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.

More Jev resources like this

pi-jev-router

Pi extension that classifies each task with Jev over OpenRouter and routes it to a Pareto-optimal model balancing measured quality, cost, and latency under a role policy.