Why give them away
Rogue Digital runs on AI agents. Eleven of them plan, build, review and sell, and I’m the only person.
Running a company like that throws up small, specific problems nobody has quite solved yet. When we solve one well enough to rely on it, there’s no good reason to keep it to ourselves.
So last weekend four of them went public on GitHub, under my name, MIT licensed. Each one does one job, and each one guards against one mistake.
guardian-check: green locally, green on deploy
The failure it exists for is familiar to anyone who ships TypeScript. The typecheck passes, you push, and the production build fails on the host. Some errors only show up in the real build: very deep type instantiation, isolatedModules errors, a server component importing something it shouldn’t.
guardian-check is a Claude Code skill. It runs your typecheck, your production build and your unit tests, then works through the failures in a loop. It stops after ten attempts and it never commits, so you still review the diff yourself.
The part I care most about is flaky tests. When a test fails once and passes on the retry, the easy thing is to move on. guardian-check records it and sends an agent after the cause, with instructions to change nothing and only propose a patch. A flake you retry away is usually a race or a cleanup gap, and it comes back.
It’s built in two layers on purpose. A plain Node script with no dependencies runs the checks and writes a verdict you can trust, in CI or anywhere else. The skill is the part that needs an AI: reading the failure and fixing it.
hushbrain: memory that keeps private notes private
Agents need memory. Ours search a shared store of notes all day: decisions, projects, facts. Some notes should never turn up in a work search.
The risk I was designing against isn’t a hacker. It’s an innocent search for “what did we decide about pricing” pulling up a private note, which then ends up in an email or a commit message.
hushbrain is an MCP memory server built around that. Every note has a sensitivity tier: normal, personal or legal. Every read leaves out personal and legal notes unless the caller explicitly asks for them. Searches return a one-line summary, never the full note, and reading a full note is a separate call that refuses a hidden one.
The rule I’d copy even if you never use the code: a note with no label is private. So is a note with a misspelt label. Forgetting to label something can only hide it, never expose it.
It’s version 0.1. It runs and its tests pass, and the roadmap in the repository says what is still to do.
anthropic-cost-by-key: which key spent it?
If you run more than one thing on the Anthropic API, sooner or later you want to know which key did the spending. The Admin API has the answer, but split across two reports. The cost report has the real money but can’t be grouped by key. The usage report can be grouped by key but only has token counts.
anthropic-cost-by-key joins the two, so each key gets its share of what was actually billed, not an estimate from a price list. Anything it can’t pin to a key is listed, not hidden, and the keys plus the unattributed line always add up to the cost report total. One command, no dependencies, read only.
whatsapp-readonly: one chat, and nothing else
I wanted scripts and AI agents to be able to look up a single WhatsApp conversation, “what did we agree about Saturday?”, without being able to read the whole of my WhatsApp. Most WhatsApp tools do the opposite: they export everything.
whatsapp-readonly reads the store the WhatsApp desktop app keeps on your own Mac, and it’s built around saying no. Every read names one chat. If a name matches more than one, it lists them and stops, so you pick. There is no dump or export command, and asking for one gets a refusal that says why.
It can’t change anything either. There’s no send, no delete and no edit anywhere in the code. It copies the store, queries the copy and deletes the copy when it’s done. It makes no network connections and has no dependencies.
The rule I’d copy: text in the store was written by other people, so treat it as untrusted. A message can’t pose as a separate line of the conversation, and hidden control characters are printed where you can see them.
It’s honest about its limits. It’s a guard rail for scripts and agents, not a lock: anyone who can run it can open the same file another way. The point is that the safe path is the default, and reading more is a deliberate act.
What they have in common
None of these is clever. Each is a guard rail for a mistake that’s easy to make when software is doing the work: shipping a build that was never built, surfacing a note that should have stayed hidden, not knowing where the usage went, reading far more than you meant to.
That’s most of what running a company on agents is. You find where it can go wrong, and you put something boring in the way.
Get them
Issues and pull requests are welcome.
We build agents and automation for other businesses too. If you’re thinking about putting agents to work in yours, book a 30 minute call and we’ll compare notes.
Next in this series: where an AI company still waits on its founder, with the numbers.