Open-source · experimental alpha0.1.0-alpha.3 ↗

Run the router.
Install the client.

iRoute is a self-hosted execution runtime for task-aware AI systems. Resolve trusted state first, route only unresolved work, and keep every execution durable and observable.

Latest release0.1.0-alpha.3

Release focus

Routing claims you can audit.

Model profiles now identify their inputs as Synthetic, Unverified, or Measured. A measured profile becomes routing-eligible only when its provider, model, timestamp, and sample count are present.

  • 01 Explicit model-profile provenance
  • 02 Atomic approval resumption
  • 03 Complete workflow usage accounting

01

One runtime, thin clients

Your applicationinstalls one SDK
iRoute APIruns in Docker
Worker + databaseexecutes durable work

The packages do not embed a second router. They serialize the public HTTP contract, handle execution operations and SSE replay, and call the server you operate.

02

Local in two commands

docker compose -f deploy/compose.sqlite.yaml up --build --wait
curl --fail http://localhost:8080/health/ready

This starts the API and execution worker with durable SQLite storage. The deterministic development gateway requires no provider key.

03

Six thin clients

Alpha.3 is live on NuGet, PyPI, Packagist, and the signed GitHub release. npm and crates.io remain installable at alpha.2 while their alpha.3 trusted-publisher connections are finalized. Maven Central remains planned; Java installs from signed source.

04

Honest by design

Alpha.3 improves the evidence behind a routing decision and the lifecycle around durable execution. It does not turn an experimental project into a production promise.

What it establishes

Durable workers, deterministic routing, circuit breaking, provider fallback, observability, and six conformance-tested client surfaces.

What it does not claim

No SLA, no production connector claim, and no assertion that built-in synthetic profile values are real provider measurements.

Read the alpha.3 release notes →