Use case · Local-first developers

Local models first, cloud only when you need it

You run models on your own hardware for privacy, cost or just because you can. AutoJev lets every coding agent use them, prefers them when they are good enough, and falls back to a cloud model when they are not.

Gateway listens on localhost only
Logs store metadata, never prompts or responses
"Prefer local models" in intelligent routes
Cloud failover before output starts

Download AutoJev Star on GitHub

Automatic failover

autojev/fast

1
Attempt 1 · openai/gpt-5-mini
HTTP 429 · Retry-After 30s
2
Cooldown
openai/gpt-5-mini · 30s
Attempt 2 · deepseek/deepseek-chat
200 · streaming

The local-first dilemma

Agents assume the cloud

Codex expects a Responses API, Claude Code expects Anthropic Messages. Your Ollama server speaks neither out of the box for those agents, so local models sit unused.

Local is not always enough

A local model handles quick edits fine, but a long refactor with dozens of tool calls may exceed its context or your patience. Then you are back to editing configs.

Privacy is all-or-nothing

Most routers are cloud services. Sending every prompt through a third party defeats the point of running models locally.

You cannot see the trade-off

Without measurements, “local is slower” or “local is fine” is a guess.

How AutoJev handles local-first

AutoJev is a desktop app, not a cloud service. Its gateway listens on 127.0.0.1 only, there is no AutoJev account and no relay: requests to Ollama stay on your machine, and requests to cloud providers go directly to them with your own keys. Request logs store metadata only — full prompts, responses and API keys are not persisted in request logs.

Every connected agent — Codex, Claude Code, Gemini CLI, OpenCode, Hermes and more — can use your Ollama models, because AutoJev converts their protocol to the Chat Completions API that Ollama serves. Routes decide when local is used: always, first by priority, or whenever a local model matches the cloud candidates on capability and tier.

Setup recipe: a local-first route with cloud backup

1
Add Ollama
Providers → Add provider → Ollama. The base URL is preset to http://127.0.0.1:11434 and no key is required. Add your local models by the names `ollama list` prints, with honest tier, capability flags and context window.
2
Add one cloud fallback
Add a cloud provider such as DeepSeek, OpenRouter or any OpenAI-compatible endpoint, and one or two models to catch what local cannot handle.
3
Create the route
Create a route `local` with load balancing. Put the Ollama model in a higher priority group than the cloud model. Requests go to the local model; if it fails with a connection error or 401/402/403/408/429/5xx before output starts, AutoJev fails over to the cloud model.
text
Route: autojev/local Strategy: Load balancing Candidates: ollama/qwen3-coder:30b priority 2 weight 1 deepseek/deepseek-chat priority 1 weight 1
4
Or let AutoJev decide
Use intelligent selection with “Prefer local models” instead: AutoJev chooses Ollama whenever its capability and quality tier are equivalent to the cloud candidates, and a stronger cloud model when the request needs it.
5
Connect agents and measure
Select `autojev/local` for your agents and connect. Run speed tests to see first-content latency and tokens/s for local vs. cloud on your hardware.

Failover from local to cloud

Before output starts, a failed candidate is skipped and the next one is tried. Each candidate is tried at most once, with a 30-second cooldown by default.

Automatic failover

autojev/fast

1
Attempt 1 · openai/gpt-5-mini
HTTP 429 · Retry-After 30s
2
Cooldown
openai/gpt-5-mini · 30s
Attempt 2 · deepseek/deepseek-chat
200 · streaming

What stays local, and what does not

Data	Where it goes
Requests to Ollama models	Your machine only
Requests to cloud models	Directly from your machine to that provider, with your key
Request logs	Local SQLite; metadata only (tokens, latency, cost, status)
Provider keys	Local ~/.autojev/autojev.db (currently unencrypted)
Jev decisions (Balanced / Lower cost / Higher quality routes only)	Routing metadata to OpenRouter, never conversation content

Direct model calls, load-balancing routes and the Lower latency preference never call Jev — they are decided entirely on your machine.

Why local-first users pick AutoJev

Free and open source (MIT license); runs on macOS, Windows and Linux.
One gateway for every agent, so local models are not limited to one tool.
Speed tests on your own hardware, scheduled every 30 minutes by default.
Usage broken down by provider, so you can see how much traffic stays local.
Session affinity keeps a conversation on the same model while it stays eligible.

Frequently asked questions

Can I use AutoJev completely offline with Ollama?

Requests to Ollama go from the local gateway to your Ollama server, so they do not need the internet. Call a single model directly, or use load-balancing or Lower latency routes, to avoid any Jev call. Update checks and cloud providers, of course, need a connection.

Does AutoJev store my prompts?

No. Request logs keep metadata such as agent, model, tokens, latency, estimated cost and status. Full prompts, responses and API keys are not persisted in request logs.

How does "Prefer local models" work?

In intelligent selection routes, AutoJev picks an Ollama model when its capability and quality tier are equivalent to the other candidates. It is not shown for the Lower latency preference, which ranks by measured speed.

What if my local model is too slow or crashes mid-task?

Failover covers errors before output starts. For speed, set a route to Lower latency and let measured first-content latency and tokens/s decide between local and cloud models.

Can I use LM Studio or vLLM instead of Ollama?

Yes, add them with the OpenAI Compatible preset. Note that "Prefer local models" specifically recognizes the Ollama provider type; for others, use priority groups.