Distributed LLM capacity control

Protect shared capacity before overload reaches the provider.

Korrx (pronounced "kor") allocates capacity across the system. Tyr enforces it locally before each LLM request begins.

Fast local rejection · Priority-aware capacity · Exact decision provenance

Overload is stopped early. Tyr rejects excess work before a provider call is created.

Authority expires safely. Tyr fails closed when its Korrx grant is no longer valid.

Every decision is traceable. Grant, revision, and controller epoch remain attached to the admission.

How it works

One control loop. Fast decisions at the edge.

Korrx coordinates shared capacity in the background while Tyr keeps the request path local and predictable.

01

Allocate

Korrx divides shared concurrency and token capacity among Tyr instances using versioned, expiring grants.

02

Decide locally

Tyr evaluates each request against the current grant without adding a control-plane call to the request path.

03

Reject before execution

When capacity is unavailable, Tyr fails fast. Rejected work never reaches the model provider.

Architecture

Focused layers, one admission decision.

The control plane distributes authority. The data plane enforces it without consulting Korrx for every request.

Architecture diagram: applications send requests through Tyr, which uses async-bulkhead-llm and async-bulkhead-ts to enforce capacity allocated by the Korrx control plane before forwarding admitted requests to LLM providers.
Korrx allocates shared capacity. Tyr enforces it locally. Rejected work does not reach the provider.

System layers

Each component has one clear job.

Distributed allocation

Korrx Control Plane

Owns capacity pools, agent registration, controller epochs, and versioned grants.

Local enforcement

Tyr Admission Controller

Makes per-request decisions, reserves in-flight capacity, and records admission provenance.

LLM capacity engine

async-bulkhead-llm

Adds token reservation, priority capacity, usage reconciliation, and admission-time revisions.

Concurrency foundation

async-bulkhead-ts

Provides the base bounded bulkhead used to keep excess concurrent work out of the system.

Project status

Integration-validated and under active development.

The core path has been verified for active grants, local overload rejection, admission provenance, fail-closed expiration, and recovery under a renewed grant.

Current priorities include workload identity, authentication, OpenTelemetry, durable audit exports, and production deployment support.

Built for systems under load

Keep important LLM traffic moving without creating more overload.

Review the system