Turn on auto-update and run the newest dira (v1.2.3)!See the releases

Local multi-agent management system

Build your own AI team

Throw a request in however you like and it still lands. Watch the whole thing live, the way you would in jira: the tickets being split up, the agents working together to finish them. Your own multi-agent system, on your own machine, and it's easy to set up.

with Claude Code · Codex

The dira board. Ticket cards sit in the three lanes Open, In progress and Done, and one of them crosses into the next lane.
Fun fact: the dira app was built with dira.

Say it and it gets done

  1. .md① Submit a request

    Write down what you want the way you would say it out loud. That's the whole job. The agents take the tedious, complicated rest.

  2. .wip② Everything moves at once

    The agent that takes your request works it out and splits it into tickets, one per piece of work. A worker with the right persona picks up each ticket, and together they finish what you asked for.

  3. .done③ That's it. Easy, right?

    You see what the workers read and how they change it, live. When one gets stuck mid-job, it asks you. Ask and answer the way you would with a person, and the feature you wanted is finished!

Finished work stays on the record

  • Once a ticket turns .done, one archive ticket follows it. An Archiving line shows up under the done card, and because a worker takes that one too you can watch how far it has got
  • What's left is one markdown file and one section at the foot of the ticket. The worker that took the archiving picks the facts out of the job just finished, writes them into .dira/ontology/ in the project folder, and adds a ## 아카이브 section to that ticket's body
  • The next session starts out knowing where that is. Where the ontology sits and how to search it goes out in every prompt a worker gets
  • The files are plain markdown. [[link]] ties them to each other, so a tool like Obsidian opens the folder whole. Move the project and the record moves with it

The more you hand over, the better the workers know this project. They start by reading who decided what yesterday, so you never say the same thing twice.

The In progress and Done lanes of the dira board. The second card in Done, a732ce19, carries a filing-cabinet icon and an «Archiving» line in its bottom row, and the card above it has no such line.
That line is a link. Press it and you land on the archive ticket, where you can see how far its worker has got.

Archiving and the ontology

You don't need an account

  • dira has no server. No sign-up, no login. Download it and open it, that is all, and nothing you make gets uploaded anywhere
  • Tickets and records are both inside the project folder. The queue is the single directory <project>/.dira and what it holds is markdown files. There is nowhere to set permissions, so being able to open that folder is the permission
  • The work does go out to the model. When a worker starts a session, the ticket body and the code it needs travel through the engine you picked and reach the model. Outside that one channel, dira takes nothing away
  • Turning usage analytics off is all it takes. We count eight things about what you press on screen. Ticket titles, ticket bodies, file paths and prompts are not carried. Turn it off in settings and nothing goes out from that moment on, and to wipe what is left, delete the one file ~/.config/dira/analytics.json and you are done
The first screen when no project is registered. Under the line saying there are none, the New project card is open, with the Name, Project folder, Integration branch and Spec document fields and a Create project button.
This is the screen the first time you open it after installing. You fill in a name and a project folder, and there is no field for an account.

Usage analytics, and how to turn them off

  • 0Engine dependencies
    bash + the python3 standard library
  • 14Workers running at once across 4 projects on one Mac
  • 4237Tickets its own queue took
    4235 done
  • 44 daysSince the first commit
    2,673 commits

As of 2026-09-11

Install

Download it, install it, done

Get it and open it. You never have to open a terminal.

  • ① .dmg — open it and drag. dira.app goes into Applications and the install is over. The build is signed and notarized, so the Mac won't stop you on first open with a warning about an app it doesn't know
  • ② Open the app for the first time and the form is already unfolded. Put in a name and a project folder, then press Create project.
  • ③ Thirty seconds later a worker starts sweeping the queue. Leave a ticket there and it gets picked up from then on
  • The engine is one of four — claude, codex, grok and agy. You pick the model along with the worker, and a name that isn't in the list you type in yourself. After that, press the Engine column on the workers screen to change it. Interject is on claude only, and the session stream is on claude and grok. The app runs on Apple Silicon Macs only
  • You can also run the engine alone, with no screen. On Linux, or when you don't need a screen, follow the route that gets the repo directly.

Full install guideCreate your first projectRunning the engine alone

Plans

Start free on your own machine

  • Free
    • Install and run the local app and engine yourself
    • P2P collaboration with your teammates Soon
    • Engine MCP Soon

    The local engine and app are free forever. dira is rooting for a multi-agent ecosystem that builders make.

  • Pro
    • Cloud projects Soon
    • dira's own cloud LLM Soon
    • Web hosting for what you build Soon
    • Cloud workers Soon
  • Enterprise
    • Enterprise-only customization Soon
    • Hooks into your internal tools Soon
  • Persona market Soon
    • We build the ecosystem together too

No sign-up, no payment. Delete it if it doesn't suit you, so install it and see.