Skip to main content
Druks finds the work, does it on the tools you already use, and brings the important decisions to you. Druks is a self-hosted or cloud platform where AI agents such as Claude Code and Codex do a company’s recurring engineering work, and a person approves each change.
  • It finds the work. A schedule, an alert, a ticket or a button starts it. Nobody writes a prompt.
  • It does the work with your tools. Claude Code or Codex, with your own Claude or OpenAI account, in a sandbox, on your real repositories. You see the cost after every step.
  • You make the call. The run stops and asks you before it changes anything: Approve, Change or Stop. It waits up to 14 days, and no agent runs while it waits.

What a run looks like

The included Software Factory app takes a ticket to a pull request:
  1. A ticket or a GitHub event starts the run.
  2. A coding agent writes the change in a sandbox and checks it.
  3. When the app needs a decision, the run stops and asks you. You see what it wants to do, why, and what it cost so far.
  4. It opens a pull request for review.
Each question, answer and cost stays in the run history.

Two ways to start

Druks Cloud and self-hosted Druks work the same way. You get the same apps and you approve in the same places. The only difference is where Druks runs.

Druks Cloud

Druks runs on our servers. Sign in and connect your repositories. Nothing to install.

Self-hosted

Druks runs in Docker on your laptop or your own server. Your AI installs it for you.

Where to go next

How Druks works

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