/what
Policy decides, not the prompt
Most routers pick a model by price or speed. Yserve first asks what is allowed to leave the machine, and only then where the work should go.
One endpoint
An OpenAI-compatible address on your own machine. Point your editor at it and choose automatic routing or local only.
Policy before anything else
Yserve works out which repository a request belongs to and applies the strictest matching rule. Anything it cannot identify is treated as proprietary.
Secrets stay home
Requests are scanned for likely credentials. A match, or content that cannot be scanned, keeps the request on the device.
A log of every decision
One line per request: the route, the reasons, the attempts and the outcome. Never the prompt or the reply.
/status
Where it stands
An early build. The routing core works and is tested; most of what makes it pleasant to use is still ahead.
- Policy routing
Routes by repository and region rules, with the most restrictive rule winning.
working - Failover
Moves to the next allowed location when a model is unreachable or slow. It never widens what policy allows.
working - Decision log
Records route, reasons and outcome for every request, without prompt or reply text.
working - Secret handling
Detects likely credentials and keeps the request local. It does not redact them yet.
detects, no redaction - Test with real local models
So far the core is tested against stand-in model servers. The run on real hardware has not happened yet.
still needed - Editor extension
A VS Code companion that starts Yserve and shows where each request went.
not built - Task-aware routing
Classifying the task, checking a local answer and escalating only when it falls short.
not built - Public release
The repository is private for now. A first release follows the real-hardware test.
not yet
Honest limits
What it doesn't do yet
- You cannot install it yet; the repository is private.
- It detects likely secrets but does not remove them from a request.
- It does not judge whether a local answer was good enough.
- There is no editor extension; setup is by hand.
/next
What comes next
- Run the phase 1 test with real local models on real hardware, and record the result either way.
- Replace pattern-only secret detection with a maintained rule set.
- Add sign-in and mutual TLS for company-hosted models.
- Scaffold the VS Code companion extension.
Get involved
Can’t send your code to a public API?
The most useful help right now is a team with that constraint, willing to try local-first routing on real work and tell us where it breaks.