Skip to main content
Druks is a self-hosted runtime for durable agent apps. These apps can cross process boundaries, wait for people, and retain an inspectable history after a restart. Your app owns the domain: its workflows, agents, prompts, models, routes, and policy. Druks owns the operating substrate around it: durable execution, Postgres-backed state, queues, gates, sandboxes, harnesses, events, settings, and the shared dashboard.

Why use Druks

An ordinary agent script assumes that its process stays alive. Real work breaks this assumption. Models take time. Sandboxes disappear. Providers retry webhooks. Deployments restart workers. A human decision can arrive hours later. Druks records completed durable operations. When execution recovers, the workflow starts its orchestration again. It reuses the recorded results at the same operation boundaries. Work that stops inside an operation can run again. Thus, external writes still require idempotency. That gives an app a practical control loop:

What you can build

Druks fits apps whose work spans durable operations, external triggers, isolated agent calls, or waits. Examples include software delivery, incident investigation, research review, approval flows, and periodic operational checks. The bundled Software Factory app coordinates coding agents from a work item to a reviewed pull request. It demonstrates the framework. GitHub policy and software-delivery behavior belong to the app, not to Druks.

Choose a path

What Druks is not

Druks is not a model SDK or a sandbox provider. It does not preserve a live agent process through a crash. It does not resume at an arbitrary Python line. It does not guarantee exactly-once external side effects. Drukbox provisions hosts. Model providers and harness subscriptions remain separate. Your app owns its domain policy. Druks is under active development. Breaking changes can occur before version 1.0. main and latest are edge builds, not stable releases.