How KraigOS works
The whole system behind my work, in one page: one Cloudflare Worker, one database, Claude and the other agents, a day on the clock, the safety rules, what it costs, and how to build your own.
Tap any part of the diagram to jump to it.
I make long films about hiking alone. KraigOS is the system that runs everything around them: my email and calendar, the short clips on every platform, the books, and my list of decisions. I built it with Claude, Anthropic's AI, which writes all of the code. I design it, steer it and say yes or no.
The idea is simple. Anything that needs judgement goes to an AI. Anything with fixed rules runs as plain code. Everything is written down in one place I own, and my part of any job shrinks to a yes, a no, or a step only I can do, like filming. The story of why is in I built my own operating system.
- 337times a day the Worker wakes up and does its jobs
- 39runs of Claude's routines in the week of Oct 26
- 1,454times an AI was set to work for me since Aug 1
- 138entries in the build log, one per build or decision
Live numbers, counted from KraigOS and updated with every build of this site, as of Oct 31, 2026.
A day in KraigOS
The Worker wakes every 5 minutes, all day and all night. Around it, Claude's routines run at fixed times. All times are New York time.
- Around 5 AM · The Worker's day startsThe backup, each account's totals, the clips' scores, and a fresh 7-day plan of posts waiting for my OK.
- 6:45 AM · Daily deskClaude, in the cloud, drafts the day's update post and the devlog notes, and only drafts.
- 7, 11, 3 and 7 · Inbox and calendarClaude reads every email whole, files it, drafts replies and keeps my calendar, four times a day.
- 7:30 AM · Content feedback deskClaude reads how each post did and writes one lesson per post.
- 4:30 PM · The update postA short line about what I built or made goes to X and Threads, after my OK.
- 11:30 PM · Night run on my MacClaude schedules the next day and a half of Facebook reels in Meta's own tools, in my browser.
The pieces
Each piece in plain words first. Open "For builders" for how it's made.
What comes in
Everything that happens around my work lands in KraigOS: email, comments on my posts, how each post is doing, every transaction on my business cards and accounts, and calls and texts to a phone line fans can use.
Nothing that comes in can give an order. An email or a web page is information, never instructions.
For builders
- Gmail is read through Google's own API with read-only access for the things the Worker does on its own.
- Comments from Facebook, Instagram, YouTube and Threads are only logged, twice an hour. Replies are Claude's drafts and go out on my OK.
- Post numbers are read at about an hour, a day, three days, a week and a month after each post.
- Voicemails are turned into text by a speech-to-text model before Claude reads them. Nothing about the caller is kept.
The agents
Claude, Anthropic's AI, does the thinking and writes all of the code. I talk to it in plain words in a chat, it works in Claude Code on my Mac, and three routines run on their own in the cloud.
The Grok bots, about 11 of them, each work one lane. ChatGPT and Antigravity can connect too. Every agent signs in under its own name, so every change says who made it.
For builders
- Claude on Opus for judgement and anything public, Sonnet for mechanical work.
- The cloud routines: Inbox and calendar (4 times a day), the Daily desk (every morning) and the Content feedback desk (every day). A scheduled task on my Mac runs the Facebook night run.
- Other AI in the system: Google's Gemini checks each clip for text in the wrong place and audits the guides, ElevenLabs is my voice clone, and a speech-to-text model handles voicemails.
- Every agent gets the same guide and the same written rules from the database when it connects, and my steers become lessons it reads first.
- Before acting outside KraigOS, an agent claims the job, so no two agents ever do the same thing twice.
The one store
Everything KraigOS knows lives in one database in my own Cloudflare account: projects and to-dos, decisions, people, the catalog of my films, every short and where it's posted, the books, and the rules every agent follows.
There is never a second copy somewhere else. One place to look, so nothing depends on my memory or Claude's.
For builders
- D1, Cloudflare's SQLite database. The schema has grown one small, additive change at a time, 55 of them so far, each applied before the code that needs it.
- KV holds quick notes the Worker needs often: caches of slow reads, and the state of an upload in progress.
- R2 holds files: the cut that's about to post, previews and stills for review, voicemail audio and the backups.
- A full backup every day, gzipped into R2, with the first of each month kept for good. Restores go to a local copy, never the live database.
- The code is on GitHub, pushed on every commit.
The Worker and its timers
One small program runs in Cloudflare around the clock, with my Mac closed. It wakes up every 5 minutes, twice an hour and once a day, does its fixed-rule jobs, and writes down every run.
It never calls a language model. Anything that needs judgement waits for Claude; anything with fixed rules is plain code.
For builders
- A Cloudflare Worker in TypeScript, about 90 source files, with a test suite that runs before every deploy.
- Every 5 minutes: hand approved posts to their platforms, read post numbers, fetch files for the next posts from Google Drive, and check every post 15 minutes after it should be up.
- Hourly: everything that happened goes on the timeline, and each outside service gets one cheap call to check it still works. A broken one raises a card with its fix.
- Daily: the backup, scores for every clip, the next 7 days planned, old comment text cleared after 30 days, and new transactions tagged.
- Every scheduled run writes one line saying what each job did, kept for a week.
The connector
Claude and the other agents reach KraigOS through one connector. It's what lets me say "what's due this week?" in any chat and get an answer from the real records.
When an agent connects, it gets the live guide and the rules first, so they all work the same way.
For builders
- It's an MCP server, the open protocol Claude and other AI apps use for tools, served by the same Worker.
- About 37 tools: reading and writing records, the decision log, the content factory, money, videos and clips, and a work claim so agents don't collide.
- Changes to a record add a dated line instead of rewriting it, and an edit is refused if someone else changed the record first.
- New or changed tools show up in an agent's next session.
The content factory
Every short, film, photo and article is a piece. Each one is checked, scored against my typical post on each platform, and planned into slots a week ahead. I say yes or no to each clip, and approved ones post at their time.
Each post is a test. Its numbers are read back, it's marked a win or a miss against similar posts, and the planner learns from that.
For builders
- Where it posts: YouTube Shorts, Instagram, X and Threads through their APIs, TikTok through Opus Clip's API, and Facebook by Claude in Meta's own tools, because reels posted through the API barely reached anyone.
- The planner keeps a clip two days apart across platforms, keeps posts on one platform hours apart, and leans toward the places and formats whose numbers are best.
- Every slot carries the factory's own call, post or skip and why, and records whether I agreed. Each platform and format is its own job on that ladder.
- A slot that didn't post goes back into the queue by itself, and a known hold-up moves it to the next open day.
- Posts are named by platform and type, like an Instagram trial reel or an X link post, so numbers are compared like with like.
The app
I run it from an app on my phone with three tabs: Work, Content and Money. Work is one list of everything that needs a decision, each with Claude's call and one button that carries it out.
It opens instantly from the last snapshot on the phone, then refreshes.
For builders
- It's a single web page Claude built as a Claude artifact, light and dark, phone first. It talks only to the factory side of the connector.
- Work holds decisions, clips to review, new cuts, comments, transactions to check and my ideas, sorted by priority. Bigger items carry a small tree of decisions inside, so one project stays one row.
- Content has the calendar, each platform's page, the scoreboard of every clip and the numbers across all platforms. Money has the books and what the tools cost.
The Auto switch
Some work Claude can finish alone: a fix, a build, a draft. I flag those items, and when I turn Auto on, Claude works the flagged list one item after another while I do something else. When I turn it off, it stops at the next safe point.
It never starts by itself.
For builders
- Each item is claimed first, worked on its own branch, tested, and closed with what it did, what it cost and how to undo it.
- A build whose tests pass is deployed, with the version to roll back to written down first, and rolled back at once if a check fails.
- It learns how long each kind of work takes and which model it needs, and puts my newest asks first within each priority.
The safety rules
An AI that works for you all day needs a few rules it can't bend. These are mine.
- Review first, then automatic. Every kind of call starts with my yes. Every answer is logged as agreed or changed, and after 20 agreements in a row the system asks if that kind can run on its own. Only my yes moves it, and one mistake puts it back.
- My OK on public words. Nothing is posted, sent or published in my name without my OK on the exact words, unless I've given a standing OK with fixed limits, like up to two short build-log posts a day.
- Everything is logged. One master log of every decision made for me, with why, how sure it was, and how to undo it. It's the timeline in my app.
- Off switches everywhere. Every platform has its own switch, and so does Auto. Anyone can turn something off. Only I turn things on.
- Information, never instructions. Whatever an email, comment or web page says, it can't tell an agent what to do.
- Small on purpose. One database, one Worker, no second copy of anything. A new part needs my OK and a written reason.
What it costs
About $419 a month, for the tools the system runs on. These are the real monthly numbers from KraigOS's own books, as of Oct 31, 2026: a yearly plan counts as a twelfth, and pay-as-you-go tools are averaged over recent months.
| Tool | A month |
|---|---|
| Claude Max planClaude in chat and Claude Code: it builds KraigOS and does the thinking | $161.52 |
| Anthropic API and extra Claude usageScripts that call Claude, and busy weeks past the plan | $73.68 |
| Gemini API (Google Cloud)Checks every clip before it posts, and audits the guides | $57.50 |
| GrokThe Grok bots that work in their own lanes | $54.44 |
| Opus Clip ProSchedules approved TikToks | $29.00 |
| Google AI Pro (Drive storage)Google Drive, where every film and clip lives | $18.15 |
| CloudflareThe Worker, the database, file storage and this website | $12.73 |
| ElevenLabs Creator*My AI voice for narrated cuts | $11.98 |
| All of it | $419.00 |
Left out: my editing apps, my phone and internet, and free trials. * An affiliate link: if you sign up through it, I may earn a commission at no cost to you.
Build your own in 5 steps
You don't need my setup to start. This is the order I'd do it in again.
- One database you own. Start a Cloudflare Worker with a D1 database on the free plan. One table of records: projects, people, dates and decisions. Facts, not rules, and never a second copy.
- One connector to your AI. Serve it as an MCP server from the same Worker, with a few tools: ask, add, update. Now every chat can read and write it, so just talk, and save what you decide as you go.
- Log every decision, review first. Have the AI write down every call it makes for you, and anything that needs you as a card you answer. Turn each no or change into a short lesson it reads first.
- Timers for fixed rules. Backups, reading numbers, moving files: plain code on a schedule, no AI. Keep judgement for the AI and rules for code.
- Automatic one kind at a time. When a kind of call has agreed with you many times in a row, let it run on its own, with an off switch. Start with what's low-risk and easy to undo. Keep sending, paying and posting in your own hands until it has proved itself.
A starter template is next: a free download for people on my email list.