AutoJev
Download

Local model gateway for AI agents

Your agents ask. AutoJev picks the model.

AutoJev gives Codex, Claude Code, Gemini CLI and your other agents one endpoint on 127.0.0.1. Add the providers you already pay for, and every request goes to the right model — pinned, load-balanced, or picked by Jev.

Download AutoJevStar on GitHub
  • Free & open source
  • macOS · Windows · Linux
  • Runs 100% locally

Route every request to the right model.

Connect an agent
Codex
Claude Code
Gemini CLI

AutoJev

autojev/quality
openrouter/anthropic/claude-sonnet
deepseek/deepseek-chat
ollama/qwen3:14b

Providers

4

Models

12

Routes

5

Agents

3

Requests today

1,284

Tokens today

18.6M

Connects the agents you use

To the models you choose

How it works

One local hop between your agents and every provider.

Agents talk to AutoJev on localhost. AutoJev picks the model, speaks the provider's protocol, retries when something fails, and writes a log line — then streams the answer back.

Your agents

  • Codex
  • Claude Code
  • Gemini CLI
  • OpenCode

AutoJev gateway

127.0.0.1:9527
  • Pick a model per request
  • Convert Chat ⇄ Responses ⇄ Messages
  • Fail over, cool down, retry
  • Log tokens, latency and cost
/v1/chat/completions/v1/responses/v1/messages/v1beta/models/*

Your providers

  • OpenAI
  • DeepSeek
  • OpenRouter
  • Ollama
  1. 01

    Add providers

    Paste a key for OpenAI, DeepSeek, OpenRouter, Ollama or any compatible endpoint — or import them from CC Switch.

  2. 02

    Create routes

    Group models into routes like autojev/fast or autojev/quality, and choose how each one decides.

  3. 03

    Connect agents

    One click writes the agent's own config, with a backup. Quit AutoJev and everything is restored.

One-click setup

Connect in one click. Undo in one click.

AutoJev detects the agents on your machine and wires each one to the models or routes you pick — in the agent's native config, nothing to hand-edit.

  • Codex, Claude Code, Gemini CLI, OpenCode, Hermes, OpenClaw, Kimi Code, Grok Build and oh-my-pi — plus custom agents.
  • Original config is backed up on connect and restored on disconnect — later edits you made are kept.
  • Give one agent several models or routes and switch inside the agent's own model picker.
  • Any other tool gets copy-ready base URL, key and model snippets in env, JSON or cURL.
How one-click setup works

Agents

+ Agent

CodexResponses

~/.codex/config.toml

Connected

Claude CodeMessages

~/.claude/settings.json

Connected

Gemini CLIGemini

~/.gemini/settings.json

Connected

OpenCodeChat

~/.config/opencode/opencode.json

Connect

Smart routing

The right model for every request.

A route is a model ID your agents call, like autojev/fast. Behind it, you decide how AutoJev chooses.

Jev, a decision model on OpenRouter, helps pick a candidate for each request from routing metadata only — never your conversation. If Jev is unavailable or unsure, AutoJev falls back to local selection and load balancing.

Decision preference

autojev/balancedIntelligent selection
  • openrouter/claude-sonnet
  • deepseek/deepseek-chat
  • openai/gpt-5-mini
  • ollama/qwen3:14b

Failover before output

Connection errors, 401/402/403/408/429 and 5xx move the request to the next model — up to 4 attempts.

Cooldown & Retry-After

A failing model sits out for 30 seconds by default, and AutoJev honors the provider's Retry-After.

Prefer local models

With the lower-latency preference, tell AutoJev to favor your Ollama models and reach for the cloud only when it has to.

Savings you can see

Set a cost baseline and AutoJev estimates what routing saved compared with always using it.

Explore smart routing

Protocol conversion

Codex speaks Responses. Your provider speaks Chat. AutoJev translates.

Codex speaks Responses. Claude Code speaks Anthropic Messages. Many providers only offer Chat Completions. AutoJev converts all six directions — streaming, tools, images and usage included.

See what converts

One endpoint, three protocols

127.0.0.1:9527/v1

Chat Completions
Responses
Messages

Chat Completions

/v1/chat/completions

—

Responses

/v1/responses

—

Messages

/v1/messages

—
ResponsesDeepSeek · Chat
MessagesOpenAI · Responses
ChatAnthropic · Messages

Speed tests

Measured, not guessed.

AutoJev measures first-token latency and tokens per second for every model on a schedule, so “lower latency” routes are based on your own network and keys.

How speed tests work

Models

▶ Speed test
ModelFirst tokenSpeed test
deepseek-chat620 ms58 t/s
claude-sonnet940 ms71 t/s
gpt-5-mini780 ms96 t/s
qwen3:14b210 ms42 t/s

Logs & usage

Every call on the record — on your machine.

See which agent called which model, through which route, how long it took and what it cost. Logs keep metadata only — never prompts, responses or keys.

Logs

TimeModelStatus
14:02:31deepseek/deepseek-chat200
14:02:18openrouter/claude-sonnet200
14:01:55ollama/qwen3:14b200
14:01:40openai/gpt-5-mini429
14:01:12deepseek/deepseek-chat200
Explore logs & usage

The details

The parts you don't see.

Local-first

The gateway listens on localhost only. Requests go straight from your machine to your providers.

Bring your own keys

No AutoJev account, no relay and no markup. You pay your providers directly.

Import in seconds

Already set up in CC Switch or Termany? Import those providers, de-duplicated by endpoint.

Debug console

Try any model or route in chat or cURL mode and inspect timing, tokens, cost and raw JSON.

Lives in the menu bar

Close the window and the gateway keeps running. Pause it, or quit and restore every agent.

Signed auto-updates

Signed builds for macOS, Windows and Linux, with verified in-app updates.

Questions, answered.

What is AutoJev?

AutoJev is a free, open-source desktop app that runs a local model gateway on 127.0.0.1:9527. Your AI agents send requests to it, and it routes each one to a model from the providers you configured.

Which agents does it work with?

One-click setup supports Codex, Claude Code, Gemini CLI, OpenCode, Hermes, OpenClaw, Kimi Code, Grok Build and oh-my-pi. You can add custom agents, and any other OpenAI- or Anthropic-compatible client can use the connection parameters AutoJev shows you.

Does AutoJev see or store my prompts?

Requests pass through the gateway on your own machine, but request logs keep metadata only — full prompts, responses and API keys are not persisted in logs. When intelligent selection calls Jev, only routing metadata is sent, not conversation content.

Where are my API keys stored?

Locally, in a SQLite database at ~/.autojev/autojev.db. Keys are currently stored unencrypted, so protect your machine account accordingly.

How much does it cost?

AutoJev is free and open source under the MIT license, including for commercial use. You pay your model providers directly with your own keys; AutoJev adds no markup.

What is Jev, and do I need it?

Jev is a decision model available on OpenRouter. With intelligent selection, AutoJev uses it — with your own OpenRouter key and routing metadata only — to judge which candidate fits a request. It's optional: direct models, load-balanced routes and the lower-latency preference never call it, and intelligent routes fall back to local selection when it's unavailable.

What happens to my agent configs if I uninstall?

AutoJev backs up each agent's config when it connects and restores it on disconnect or when you choose “Quit AutoJev (restore agents)”, keeping any non-conflicting changes you made in the meantime.

Which platforms are supported?

macOS (Apple Silicon and Intel, signed and notarized), Windows x64 and Linux x64 (AppImage and .deb).

Give every agent the right model.

Install AutoJev, add the providers you already pay for, and connect your agents in one click. Everything stays on your machine.

Download AutoJevStar on GitHub

Free & open source · macOS · Windows · Linux