Inspiration

I was inspired by Peter Steinberger and OpenClaw. I wanted to make setting up a personal agent as easy as signing up for a newsletter.

My test has always been: easy enough for my mom and dad, but powerful enough for me. Safe, secure, reliable, and actually useful. Everyone should be able to have their own agent, free for life. That's the dream.

I am a retired CTO. I came out of retirement to build this. I want to be a founder but still live like I am retired. Agents are helping me do that.

What it does

Octogee gives each person an AI partner, a private workspace, chat, projects, repositories, Canvas pages, tasks, sessions, automations, previews, a Builder Shell, and the OpenClaw Dashboard.

It is a real platform running across Firebase, Mattermost, Gitea, Google Cloud, Kubernetes, and isolated OpenClaw gateway pods. It is not a chatbot wrapped in a landing page, and the demo is not a one-process mockup. A real Octogee session crosses all of those systems.

What was already there

Octogee existed before Build Week. The Mattermost and Gitea foundations, original Builder Shell, isolated gateway architecture, Playbooks, Chief-of-Staff model, and Free-for-Life plan all predate the event.

I am not claiming those as Build Week work. The README has the cutoff, commits, pull requests, tests, and Codex-session receipts for what changed after July 13.

What I built during Build Week

Then I went a little crazy.

  • Native OpenClaw chat with streaming Markdown, better rendering, completed-work rollups, and queue/reconnect reliability.
  • Automations, Canvas and Activity improvements, workspace Tasks, project Sessions, model controls, and agents that create real Octogee projects instead of fake demo artifacts.
  • A secure in-app OpenClaw Dashboard and a tenant-scoped, read-only Ops Agent that can help without wandering through customer secrets.
  • Named, expiring, auditable, and revocable permissions for sharing the Dashboard and Builder Shell.
  • Google OAuth, Drive Picker, Calendar access, Agent Import, and much safer customer deletion.
  • A simpler release and gateway rollout system using immutable releases and native Kubernetes behavior.
  • A completely isolated Build Week environment so judges can use the real product without touching production customers.

The measured window contains 585 mainline commits and 171 merged pull requests. The retained tree added more than 72,000 lines of first-party source and 60,000 lines of tests. Those numbers are receipts, not a claim that lines of code equal value.

How I built it

Octogee is 100% vibe coded. I know what that means. I am a CTO. And I still think it makes sense.

I built it with OpenClaw, my agent Bender.Echo, and GPT-5.6 in Codex. Codex helped turn product decisions into tests and implementation, coordinate changes across the PWA, APIs, gateway, and infrastructure, find integration bugs, review patches, merge fixes, deploy them, and prove that the result worked.

I did not hand over the product decisions. I decided what Octogee should be, where the trust boundaries belong, what was good enough to ship, and what needed to be deleted. Codex made it possible for one person to execute at a speed that would otherwise require a team.

The final representative Codex session ran for about 48 hours. Actually working towards that goal for 1d 11h 3m 25. It set up the isolated environment, ran the complete demo using disposable accounts, found bugs, fixed them, merged and deployed the fixes, and ran it again. That session gets as much credit as anybody.

The repository-bound audit found 262 Codex task transcripts and 9.597 billion model tokens processed, with a 96.83% input-cache hit rate. That is nuts. It is model traffic, including repeated cached context—not 9.597 billion unique words or an API-cost estimate. Frankly, I think the Codex team will appreciate that cache hit rate.

The failed experiment

We spent part of the week extending a nine-phase gateway roll, freeze, revision, and tombstone system. Then we admitted it was the wrong design and deleted it.

Two replacement PRs recorded 6,683 additions and 53,673 deletions—about eight lines removed for every replacement line. Kubernetes already knew how to do the hard part. The simpler system is easier to understand and easier to trust.

That may be my favorite Build Week receipt. Agents can help you build the wrong thing incredibly fast. Good engineering still means knowing when to stop and delete it.

Challenges and lessons

Agents are creative. So creative they can gaslight you.

The hard part was not producing code. It was preserving clear boundaries across authentication, chat, source control, cloud infrastructure, and isolated agent runtimes while moving this fast.

Tests, explicit failures, reviews, immutable releases, and disposable test accounts kept us honest. I learned that with agents, the bottleneck moves from typing code to judgment: deciding what matters, designing the environment, testing the boundaries, and recognizing complexity that should not survive.

What I am proud of

I am a solo operator. I plan on staying that way, so the platform has to run largely lights-out. Two fixes from this week were also merged upstream into OpenClaw.

But the thing I care about most is simpler: a new person can create a free account and get a real, isolated AI partner without needing to understand any of the machinery underneath it.

Try it

Open the isolated Build Week environment in a private or incognito window. Create a free account using email and password. No payment information is required.

Judges who want the prepared Launch Room walkthrough can contact me through DevPost for dedicated credentials.

What's next

Customers. More than just my mom and dad. They were the first two, but hopefully not the last.

Let's be honest: there is no way I could have built this much without all the resets from Tibo, the amazing capabilities of Codex and GPT-5.6, and the inspiration from Peter to come out of retirement and change the world.

Thanks. It's been a blast.

Built With

Share this project:

Updates