Only what matters reaches you
A decision that needs your call waits under Needs you. What OKRA is working on stays out of the way, and finished chats fold away under Quiet.
Set your OKRs. Let an army of agents deliver.
One main chat runs Claude Code, Codex and more for you. It only asks when it matters.
Leave your email. We’ll invite you to the community and let you know when it’s ready for you to try.
How it works
You say what you want. The main chat does the coordinating. You step in only where your call matters.
-
Tell the main chat what you want
Each project has one main chat. Ask in plain words, the way you would brief a colleague: a feature, a fix, a whole batch of work.
-
It plans and runs the agents
It turns your ask into tasks and gives each one to a coding agent, such as Claude Code, Codex or Antigravity, in its own git worktree. When an agent has a question, the main chat answers it for you.
-
You see only what needs you
Decisions that need your call wait in one place. The rest is reviewed and landed on main with your authority, and every landing keeps its history.
What the main chat handles for you
Real screens from one project on Nov 4: the one that built 17 games.
Big asks become epics
An epic groups the tasks. Each task is its own thread, working in its own git worktree, and the pieces run side by side, up to ten at once here.
UX walks
Ask for a walk and an agent uses your app in a real browser, on desktop or phone, as a casual gamer or someone in a hurry. It takes a screenshot at every step and notes what it finds there.
Watch it live, or replay it later.
One inbox for findings
What the walks find lands in one place, with a screenshot and the walk it came from. Fix one, or select several and fix them together in one task.
Every landing is a version
Each landing on main is a version with its history. Open any of them to see what changed, and in which files.
Focus
Write down what matters as OKRs: an objective, plus a few measurable key results. Agents work on what matters first. OKRA can draft them; you adopt them.
Your engines, your choice
Claude Code, Codex and Antigravity today, and more as they plug in. Choose the engine and model for each role, and have a different engine review or check the work.
Checks before main
Every task shows where it is: planned, built, checked, on main. A check runs your project’s tests on a disposable copy of the change.
Why it’s built this way
OKRA runs where your code already is, with the engines you already use.
It runs on your machine
No OKRA server, no account and no telemetry. OKRA works next to your repo and keeps its record there.
It uses the coding CLIs you already pay for
Claude Code, Codex, Antigravity: each runs under your own account. OKRA is the conductor, not another model.
It never pushes
Landing is a local merge. Pushing stays with you.
Every task gets its own worktree
Agents work in separate git worktrees of your repo, so tasks running side by side stay apart.
A second engine checks the first
A different engine catches what the first one misses. OKRA started there: Claude sends its plan or its diff, and Codex reviews it.
Built by its own agents
6,584 commits in OKRA’s own repo since Jun 25, most of them written by agents OKRA was running. 3,522 in October alone.
Before its own repoIn its own repo, where OKRA builds itselfNovember, projected at the Nov 1–4 pace
Show the numbers
| Month, 2026 | Before its own repo | OKRA's own repo |
|---|---|---|
| March | 6 | – |
| April | 160 | – |
| May | 88 | – |
| June | 145 | 110 |
| July | – | 388 |
| August | – | 795 |
| September | – | 1,333 |
| October | – | 3,522 |
| November 1–4 | – | 436 |
| Total | 399 | 6,584 |
| November, projected at the Nov 1–4 pace | – | ~3,270 |
- v1A task board for agents, plus OKRs to keep them focused. That’s the name: OKRs for Agents.
- v2Rebuilt from scratch. Each task became a chat, working in its own git worktree.
- v3One main chat runs the work, and only bothers you when it matters.
5 asks started 17 games.
17.8 agent-hours of work in 6.7 hours of clock time, polishing included, with up to 12 agents at once.
5 games in 72 minutes, from one prompt.
Chess, Checkers, Mancala, Four in a Row and Reversi. The only other message: you should be driving all that
.
Before you start
What you need
- Linux, or Windows with WSL2 (macOS support is on the way).
- A coding CLI signed in from a terminal: Claude Code or Codex runs the main chat, and Antigravity can take on tasks and checks too.
- A real project in git that you’re allowed to run agents on.
Privacy, in three lines
- OKRA runs entirely on your machine: no OKRA server, no account, no telemetry.
- Your code goes only to the engines you already use, under your own accounts.
- The pilot form asks for your email and a few answers, and nothing else.
Questions
- Do I have to watch it?
- No, and that’s the point. The main chat answers its workers, reviews their work and lands it. It comes to you when something needs your call.
- Which engines does it drive?
- Today: Claude Code (
claude), Codex (codex) and Antigravity (agy), installed and signed in from a terminal, on your own subscriptions. The main chat runs on Claude Code or Codex. Engines plug in through one adapter, so others, OpenCode for example, can be added later. - What does it cost?
- Nothing for now. The agents run on your own subscriptions, and OKRA runs several at once, so it uses your quota faster than a single chat. Pricing for later is undecided.
- Where do my code and data go?
- Only to the engines you already use, under your own accounts. OKRA works in separate git worktrees, lands with a local merge and never pushes. Use a fresh clone, and keep production keys out of the shell you start it from.
- How do I install it?
- When it’s ready for you, you install it yourself from a one-page guide, with help in the community if you need it.
- Is it open source?
- Not decided. For now it’s a private checkout, and what early users show is part of deciding.
- Why not just Claude Code subagents?
- That’s where it started, on Mar 28. Subagents are great inside one session. OKRA is for the work that outlives one: hours or days, a different engine checking the work, and one chat that keeps track of all of it.
- What does the name mean?
- OKRs for Agents. OKRs are Objectives and Key Results: a goal, plus a few measurable results that show you’re getting there. In OKRA, that’s Focus.