konduktor

where humans and agents build products together

view install.sh | full installation docs

Not just an orchestrator, really

In some ways, konduktor is an agent orchestrator.

Set it up on your machine or a VPS, fire off tasks, and get PRs back.

But konduktor's goal isn't to enable us to generate AI slop faster, but rather to explore how humans and agents can build high-quality products together.

This sounds like BS, but I mean it.

Here are konduktor's driving principles:

1. Human-in-the-loop

konduktor let's you fire off background agents and even schedule agents to run throughout the day or while you sleep.

But agents follow your workflow, ask you questions asynchronously before proceeding with certain tasks, and you have full observability + tooling to review their work efficiently.

2. Spec is king

Agents are getting better at taking loose requirements and building reasonable implementations, but not only do great specs help them, they help you.

The more you think deeply about challenges and product direction, the better your product will be. konduktor encourages you to think and plan, not just do.

3. Opinionated but configurable

konduktor provides some opinionated primitives that help structure your work, like a task board and a projects feature to manage complex work with a lot of moving parts.

These let you get moving fast, but being open-source and running on your machine means you can always peak under the hood and configure things the way you want them.

What does that mean in practice?

In practice, those principles translate to the following features today:

• Background agents: Create a task, run in the background, get a PR back. Resume the session, inspect logs, add metadata, etc. Full end-to-end experience.
• Scheduled agents: Create agents that run throughout the day or overnight. Mine currently challenge my specs, hunt bugs, do first-pass reviews, and come up with ideas.
• Projects: Manage work for complex features with a spec and defined project phases. Tasks in a phase can be completed in parallel (by you, agents, or a combination of both) and phases get unblocked when all tasks from the previous phase are complete. Scheduled agents are great at managing projects and starting work autonomously once tasks are unblocked.
• Asynchronous questions: The questions API let's agents ask you clarifying questions about your tasks, forcing you to think and spec things better, and enabling them to automatically start work once they have enough context. The questions experience is really nice both in the UI and the CLI, and I start my day by answering questions.
• Kanban board: Classic feature from standard orchestrators. Oversee work from a Kanban board that both you and your agents can manage.
• Product docs: Write your vision, plans, PRDs, etc. inside konduktor. Choose what's just for you and what's required reading for your agents.
• CLI: Anything in konduktor can be managed from the CLI, making it so that you don't need to leave the terminal if you don't want to, and that whatever coding agent you use can leverage konduktor as a source of truth and easily coordinate work. Whenever I'm happy with a spec after being challenged by Claude Code, it creates a project for me with well-defined phases and fires off the first tasks.
• GitHub integration: Merged PRs get moved to "Done" automatically. Merge conflicts get resolved as soon as they occur.

Where should I run it?

konduktor runs on a machine you own, be it your laptop or a VPS. You're strongly recommended to deploy this on a remote machine behind a domain so that you can benefit from GitHub webhooks, run agents overnight (without any hassle), and let the agents operate more freely away from your personal stuff.

Get started

See the full installation docs.

Demo