Does auto mode keep GPT-4o quality at a fraction of the price?

No. Simple queries route cheaply but complex reasoning is meant to stay on premium models at full price. The router picks the cheapest capable provider rather than forcing everything downward.

What is the real difference between auto and jev-auto?

Auto mode is a near-zero-overhead heuristic path while jev-auto uses a calibrated ML decision head with slightly higher warm cost. Both sit behind the same OpenAI-compatible interface.

Do my prompts pass through a hosted service that takes a cut?

It runs on your own machine or container with no intermediary cut and no shared key requirement. Prompts stay on your infrastructure apart from the providers you route to.

What happens when a provider is down or I cannot host locally?

It can fail over to other providers and cache semantically across turns, but it cannot run without a local service and it is the wrong pick when every request must be exactly GPT-4o.

What are the actual per-query savings it claims?

Its own table prices four cases against GPT-4o: What is 2+2 at $0.03 routed to Groq at $0.0001 for 99.7% saved, Explain quantum to Mistral at $0.0002 for 99.3%, Write a Python function to DeepSeek at $0.002 for 96%, and complex reasoning left on GPT-4o at $0.15 for 0%. That last row is the design: the router is meant to decline the downgrade, not apply it uniformly.

How does a request actually flow through it?

Three stages: a semantic step asking whether this is a duplicate for a free cache hit, then the router deciding simple question or complex reasoning, then the provider. Availability is checked at runtime rather than hardcoded, across 80+ providers in free, budget, mid and premium tiers, from Ollama and Llama.cpp up through GPT-4o and Claude Sonnet.

What can I try before committing to it?

npx a3m-router serve starts the OpenAI-compatible server on port 8787 and needs no API key shared, and npx a3m-router route "prompt" previews the routing decision without serving traffic. npx a3m-router health reports live provider availability and npx a3m-router benchmark runs a local quality benchmark. Docker is one command, and there is also a PyPI package if Python is your default.

CommunityJev Agents & MCP

A3m Router

Self-hosted OpenAI-compatible router on port 8787 that sends each prompt to the cheapest capable model across free, budget, mid and premium tiers, with heuristic auto and Jev-backed routing modes.

See highlights

Open github.com

Stars: 16
Language: TypeScript
License: MIT
Published by: Community
Author: Das-rebel
Last verified: Oct 2026

Das-rebel/a3m-router

User repository on GitHub

View on GitHub

⚡ Adaptive multi-model LLM router — 80+ providers, Jev System One single-pass routing (model=jev-auto), pheromone-trail failover, parallel ensemble merge. npm: adaptive-memory-multi-model-router

Forks: 5
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

GPT-4o costs $0.03/run. A3M routes the same request to Groq/Mistral for $0.0001. Most requests don't need GPT-4o. A simple question costs the same as a complex one. A3M analyzes your prompt and routes to the cheapest provider that can answer it correctly. model="auto" — Heuristic router, 0.4ms overhead, zero extra cost. model="jev-auto" — ML decision head, calibrated probabilities, 2ms warm. Provider availability checked at runtime — no hardcoded uptimes. OpenRouter takes a cut. A3M runs on your machine.

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

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Routing

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Drop-in OpenAI base-URL replacement with model auto for cheapest-capable routing.
Heuristic and ML-backed routing paths with semantic duplicate checks before provider choice.
Tiered provider catalog from local runtimes to premium models with runtime availability checks.
Parallel ensembles, uncertainty estimation, memory windows and CLI health and benchmark tools.

Watch out

Savings illustrations assume simple prompts route correctly and complex reasoning still pays premium rates. Availability is checked at runtime and local-service or Docker hosting is on you. 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

duet-agent

Agent harness for work that outlives one chat, with resumable turns, routed state machines, and observational memory across sessions and processes.