October was the busiest building month I've had in a long time. A team of Grok Bots did most of the typing, and I did the steering, the testing and the saying yes or no.
Looking back at the month, most of what I shipped circles the same question: how do you let AI agents do real work without handing them the keys to everything? Here's the rundown.
Ideabot went live
I already wrote up how I built Ideabot in a day, so I'll keep this short. Ideabot is a daily record of what people actually do with AI agents. Every morning it collects posts from public sources, pulls out the use case, the reported result and a prompt you can copy, then publishes a ranked top 5 and a newsletter. It went from idea to live site on October 11.
The rule I care most about there is honesty. If a post doesn't report an outcome, the card says so. If a prompt wasn't quoted from the source, it's labeled as an AI-written example. That same instinct runs through everything else on this list.
BlindPass: agents use my keys, they never see them
If you build software with AI coding agents, you know the problem. Your .env file sits in the project folder, the agent can read the project folder, and sooner or later a live API key ends up pasted into a chat. Apple Passwords is great for logins, but it isn't built for API keys, tokens, .env variables and recovery codes.
So I built BlindPass, a Mac app that keeps those secrets out of files. Version 1.1.0 is out now. It's free, open source under the MIT License, and about a 5 MB download.
What it does:
- Keys and tokens in one place, with expiry dates and warnings on live keys. Press ⌘N, paste something secret, and the value turns into dots right away while BlindPass suggests where it belongs. It recognizes common services like Stripe, OpenAI, GitHub, Supabase and AWS, and it tells you if you're saving something you already have.
- Projects. Drop a project folder on it and it reads the
.envfiles, sorts them into Local, Preview and Production, and never changes them. It warns you about files git could commit, and it uses.env.exampleto show what's missing. - Recovery codes that cross themselves off. Copy a backup code and it's marked used, so you never try the same one twice.
- Health. One view for codes running out, keys about to expire, recovery keys you never printed, and keys left sitting in files.
The part I'm proudest of is how it works with AI agents. Claude Code, Cursor and Codex can start your app through blindpass run. You approve each command with Touch ID, the variables go to the process, and the actual values never reach the chat. Connected agents can list key names and ask for what's missing, but live keys stay out. And when an agent needs a key BlindPass doesn't have, BlindPass opens with the name filled in so you paste it there, not into the conversation.
Everything stays on your Mac: one encrypted file, opened through the Secure Enclave with Touch ID, with a printed Recovery Key as the backup. No account, no server, no sync, no analytics in the app. It's native Swift and Apple frameworks only.
musebot.rocks: a safety rating for every Muse add-on
Meta launched Muse, its personal AI agent, and people quickly started sharing custom connectors, skills and setup prompts for it. Muse has solid guard rails of its own, but Meta doesn't review custom connectors, and a skill you paste in can still tell Muse to auto-approve things, ask for secrets in chat or carry prompt-injection text.
So on October 30 I launched musebot.rocks, an independent directory of public Muse add-ons where every listing gets a safety rating. As I write this, it lists 182 add-ons, and 6 more are held for review.
Each add-on gets a risk score from 0 to 100 in two passes. First, pattern checks look for things like bypassing Muse's approvals, moving data somewhere else, asking for secrets in chat, hidden or encoded text, or the ability to move money. Second, an AI reviewer reads the same text, treating it as untrusted data, and a concern only counts if the reviewer can quote the exact text behind it. High-risk items are never listed and never returned to agents; they wait for a security review and then a human decision. The full method is on the How we rate page.
It's built bot-first, too. Agents can search the directory over MCP, and there's a plain index.json and llms.txt for anything that wants to read it directly. Every listing links back to its original source. It runs on Cloudflare, and every change goes to a hidden staging site before it goes live.
One honest note: when I first asked for this site, it got built as a directory for the wrong agent platform. It was fixed fast, but the lesson stuck. Now every new build starts by confirming exactly what it's for.
A new title: CTO of MediKin Health
Not everything this month was a side project. I'm now CTO of MediKin Health, a startup that matches Emergency Department physicians with the right open job. I build the website, the app and the backend, and I've added it to my work history. More on that as there's more to share.
The thread
Put the month together and the pattern is pretty clear. I'm all in on AI agents, and I let them do a lot. But the stuff I keep building is about the boundaries around them:
- Ideabot labels what's reported and what's AI-written.
- BlindPass lets agents use secrets without ever seeing them.
- musebot rates add-ons before you let them anywhere near your agent.
My bots work the same way. They build, test and deploy, but nothing goes public, gets posted or costs money without my yes. This post is no exception: a bot drafted it, put it on a preview site, and waited for me to approve it.
Agents are getting more capable every week. The tools that make them trustworthy are the ones I want to build.
If any of this is useful to you, grab BlindPass or check an add-on on musebot.rocks before you paste it into Muse. See you next month.