Question 1

Does it run the model, or just route to one?

Accepted Answer

It only routes. The Rust daemon does protocol translation, routing and forwarding, and the README is explicit that inference happens at the upstream service and that the gateway loads no model weights. The cost of that split is visibility: request data is forwarded to whichever upstream the entry picked, so local mode can still send data off the machine, and a cloud deployment adds the gateway server to the path.

Question 2

Can I point it at any OpenAI-compatible endpoint I like?

Accepted Answer

Not yet. The built-in adapters are Vercel, Laya, TypeSafe SystemOne and OpenRouter, with OpenRouter converted through structured JSON Chat Completions, and the README states this does not amount to offering an ordinary OpenAI or Anthropic entry. TypeSafe is the flexible one: it takes the official API by default with model jev-latest, or any local service speaking the same /v1/systemone shape.

Question 3

Where do I edit config, and can I read my provider keys back?

Accepted Answer

Config lives in providers.toml by default, or wherever JEV_SWITCH_CONFIG points. It is read and imported into SQLite on first start, and after that the database is the authoritative source, so an external edit shows up as drift until you import it through the console. Keys never come back out: the daemon holds them and the admin API returns masked status only.

Question 4

How much of this has actually been verified?

Accepted Answer

Split, and the README says so. Rust unit and integration tests, ts-rs generated types, UI tests and type checks, Windows icon resources and single-instance isolation are automated. The desktop WebView, tray, stable-signed APK, Android service, tile and notifications, OpenRouter live calls and cross-hardware performance are listed as still needing manual acceptance. The playground has run against a local upstream with HTTP 200 and request IDs aligned to the SQLite route trace.

Question 5

What do the local and cloud modes actually change?

Accepted Answer

Local is the default, listens on loopback and needs no call token; cloud listens publicly on 0.0.0.0, requires one, and puts administrative actions on a separate admin session. The default port is 11435 and the listen address is explicitly configurable. Both modes can still route to a local, LAN or cloud upstream.

Question 6

Is there a headless way to drive it?

Accepted Answer

Yes. jev-switch-cli goes through the daemon HTTP API and offers status, models, invoke, routes and events with JSON or table output, aimed at Termux, SSH and scripts. It deliberately holds no routing logic, no SQLite access and no provider keys of its own; the TUI and full Android interaction are still listed as later work.

Question 7

How does a route actually pick a provider, and what gets stored?

Accepted Answer

The editable DAG supports multi-hop and branching between public entries, routing nodes and provider model ports, with candidate priority and failure handling configured per route; a direct connection is just the simplest case. SQLite history keeps request ID, entry, route, status, latency, tokens and upstream count, plus the route_trace.attempts chain of real failures, retries and successes, but never the raw request body or the model answers.

Jev Switch

Highlights

Questions about this entry

More Jev resources like this

From the Jev community