Feature · Smart routing

An AI model router built for coding agents

Give every agent one model ID — a route like autojev/fast — and let AutoJev decide which real model answers each request: a weighted load-balancing pool or an intelligent choice guided by your decision preference. To pin one model, call it directly as <provider>/<model_id>.

Routes exposed as autojev/<route-id>
Load balancing or intelligent selection
Lower cost, higher quality, lower latency or balanced
Only routing metadata goes to Jev

Download AutoJev Star on GitHub

Routes

Create reusable routes and bind your agents to a route ID.

+ Add route

fast

autojev/fast

quality

autojev/quality

cheap

autojev/cheap

balanced

autojev/balanced

autojev/lb

Why LLM routing for agents needs to live on your machine

Coding agents send hundreds of requests per session. Some are a one-line rename, some are a security review across twenty files. Pointing all of them at one premium model wastes money; pointing them at one cheap model wastes your time.

AutoJev sits between your agents and your providers as a local gateway at http://127.0.0.1:9527/v1. Instead of a provider model name, an agent asks for a route such as autojev/fast, autojev/quality or autojev/cheap. The route decides which configured model actually handles the request — and you can change that decision in the app without touching the agent again.

Three ways to pick a model

Single model

Not a route strategy: point an agent at a single model ID in the <provider>/<model_id> form, or build a load-balancing route with one candidate. Every request goes to that model. Predictable, and still logged and priced.

Load balancing

Add several candidates with a priority and a weight. The highest priority group serves traffic first, shared by weight. Failed requests move on to the remaining candidates.

Intelligent selection

AutoJev classifies the latest task, filters candidates by capability, and picks one according to your decision preference — measured locally for speed, priced locally for cost (with Jev judging the needed tier when configured), judged by the Jev decision model for quality and balance.

Four decision preferences

Each intelligent route has its own preference, so autojev/cheap and autojev/quality can behave differently against the same pool of models.

Preference	How the model is chosen	Calls Jev?
Balance quality, cost and speed	Jev weighs task features against candidate prices, tiers and measured performance.	Yes
Lower cost	Compares estimated cost per request among candidates that meet the capability and tier requirements, and picks the cheapest. Unknown prices are only a fallback.	Optional — Jev can suggest the needed tier; the price comparison is local
Higher quality	Jev is instructed to favour the strongest suitable candidate for the task.	Yes
Lower latency	Ranks candidates by measured first-content latency, output speed and success rate from speed tests.	No — decided locally

One route ID, any agent

Routes are plain model IDs on the gateway. Any OpenAI-compatible client can use them — this is the request AutoJev's connection dialog generates.

Call a route through the local gateway

curl http://127.0.0.1:9527/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer autojev-local-byok' \
  -d '{"model":"autojev/fast","messages":[{"role":"user","content":"Hello"}]}'

What the Jev decision model sees — and what it never sees

Intelligent selection uses OpenRouter's Jev decision model with your own OpenRouter key, configured once in Settings → Gateway & routing → Decision model. AutoJev sends routing metadata, not your conversation.

Derived task features: endpoint, estimated context tokens, whether tools or images are needed, a complexity score and signals
For each candidate: name, configured prices, tier, tool/vision/reasoning flags, context window and recent speed measurements
Precomputed cost estimates for this request and your decision preference
Never sent: prompt text, file contents, tool results or API keys

Guardrails that keep routing predictable

Confidence fallback

Jev returns a candidate and a confidence between 0 and 1. Below 0.5 — or if Jev is unreachable, unconfigured or names an unknown candidate — AutoJev falls back to local selection and load balancing among eligible candidates. The log records why.

Session affinity

Once a session lands on a model, follow-up turns stay there while it is eligible. New sessions get the newest decision, so an agent doesn't hop models mid-task.

Prefer local models

Tick "Prefer local models" and AutoJev chooses an Ollama model when its capability and quality tier are equivalent to a cloud candidate.

Cost baseline

Pick a baseline model and AutoJev estimates savings by comparing its cost with the selected model's cost for the same request. It never affects selection or billing.

Set up your first smart route

1
Add providers and models
Add an API endpoint and key under Providers, then add models with their upstream API type, prices and tier.
2
Create a route
Under Routes, choose a route ID such as fast, pick Load balancing or Intelligent selection, and set the model scope — all enabled models or a selected list.
3
Configure Jev (optional)
For balanced or quality preferences, enter an OpenRouter key in Settings → Gateway & routing → Decision model and click Test Jev.
4
Point your agents at the route
Select autojev/fast under Agents and connect, or paste it as the model ID in any client. Check Logs to see which model answered and why.

Frequently asked questions

What is an AI model router?

A model router decides which large language model handles each request. AutoJev does this locally: your agent asks for a route ID, and the route picks a configured model by weighted load balancing or intelligent selection. To pin one model, the agent can call <provider>/<model_id> directly.

Does intelligent selection send my prompts to OpenRouter?

No. Jev receives derived task features and candidate metadata such as prices, tiers and capability flags. Conversation text is not sent. The model request itself goes directly from your machine to the provider you chose.

Do I need an OpenRouter key to use routing?

Only for the full intelligent experience. Load balancing, direct model calls and the lower-latency preference never call Jev; lower cost may ask Jev for the needed tier but compares prices locally. If Jev is not configured, intelligent routes fall back to local selection instead of failing.

Will the agent switch models in the middle of a task?

Not while the chosen model stays eligible. Session affinity keeps a session on its model; failover only moves on if the request fails before any output has started.

Can I route between local and cloud models?

Yes. Ollama models sit in the same candidate pool as cloud models. Turn on "Prefer local models" to favour Ollama when capability and quality tiers are equivalent.

Keep exploring

FeaturesModel speed testsMeasure first-content latency and output tokens per second on your own network, on a schedule, and route by the results.FeaturesFailover & load balancingWeighted pools, priority groups and sticky sessions, with automatic failover on 429s and 5xx errors and Retry-After cooldowns.Use casesFor cutting agent costsRoute routine turns to cheaper models, set a cost baseline, and see estimated spend by agent and model. No markup.FeaturesOne-click agent setupConnect Codex, Claude Code, Gemini CLI, OpenCode and more. AutoJev edits each native config, backs it up and restores it.CompareAutoJev vs OpenRouterA local router across all your providers versus a hosted API with hundreds of models. AutoJev uses OpenRouter as a provider.

Updated 2026-10-14

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.