Vigilant_Mesh_Personal

I'm a single dad with two kids. My oldest just hit the age where the questions stop being about cartoons and start being about everything else. One night I watched him type a question into a chatbot and casually included our government names. Not because he's careless. He's nine. Nobody had ever told him that was a thing you don't do.

So I went looking for parental control tools, and honestly they made me uncomfortable in the opposite direction. Most of them upload a kid's entire browsing life to somebody else's cloud. The dashboards were things I couldn't explain, with black box "AI monitoring" I wouldn't be able to justify to my own son if he asked what it did. I don't want to spy on my kids. I want to protect them in a way I can look them in the eye and explain.

That became VigilantMesh Personal. One small controller runs in our house. The kids' devices pair to it with a one-time code that expires in fifteen minutes. Every decision it makes is one I can read on a dashboard: allow, warn, or block, with the reason attached. Full prompt retention is off by default because my kids' conversations are theirs. I get the safety signal (something sensitive was blocked) without the transcript. Protection is local-first, so if the controller goes down, the rules cached on each device keep working.

Under the hood, it's Python with a local SQLite store, signed endpoint communication with replay protection, encrypted storage for anything sensitive, LAN discovery that can announce a controller but never enroll a device, and a browser extension that only talks to a loopback endpoint on the same machine.

The part I'm proudest of is the GPT-5.6 Safety Plan Studio. I'm a dad, not a policy engineer. Describing what I want is easy ("calmer school nights, and a fair way for him to ask when something's blocked"). Translating that into categories and schedules is the annoying part, and that's what GPT-5.6 does. It turns my plain-language goals into a structured, age-aware plan, but it never touches enforcement. It proposes. I review the diff, every field of it, and nothing changes until I apply it. The request is stripped down to my goals (after local redaction), the age range, and the current rule: no browsing history, no prompt content, none of my family's database.

Codex is the reason this is a tested product instead of a weekend sketch. It traced the multi-process architecture, made coordinated changes across the controller, endpoint, extension, and packaging, and caught two real regressions in live QA that my happy-path testing missed. Both have regression tests now. The thing I took away from working this way: AI is genuinely useful here as a planning assistant, and dangerous as an authority. The product is built around that line.

The hardest problems were the honest ones. Giving a parent useful visibility without building a surveillance tool, securing device communication without demanding a cloud account, and writing docs that admit what the product can't enforce yet. Windows packaging and a macOS endpoint DMG workflow are done, and the remaining signing, notarization, and OS-level enforcement work is documented instead of hand-waved.

I never wanted family safety to be invisible. I wanted it to be understandable, consent-based, and practical. Something I can explain to my kids, because I already have.

What's next for VigilantMesh Personal Sign in without secrets. There's now a path where Safety Plan Studio runs through a parent's existing Codex sign-in (the ChatGPT OAuth stays inside OpenAI's own tool) instead of pasting an API key. Next step is making that the default, zero-friction option. A store-published extension, replacing the load-unpacked developer flow with a normal one-prompt install. Real signing and notarization. The Windows executables and macOS DMG build today; shipping them signed is the next trust milestone. OS-level enforcement adapters for macOS and Linux, then an Android endpoint. The browser extension is the honest current boundary, and I say so in the docs. Growing with my kids. Age bands already exist. I want plans that loosen on a schedule as kids get older, and conversation prompts that treat independence as the goal instead of a threat.

Built With

api
browser
codex
controls
cryptography
css
cybersecurity
device
extension
family
gpt-5.6
html
javascript
local-first
macos
manifest
openai
parental
privacy
pyinstaller
python
safety
security
sqlite
v3

Submitted to

OpenAI Build Week

Created by

I'm not a security engineer by trade. I'm a dad who can code, and this was a problem I couldn't buy my way out of, so Codex ended up being the other half of the build.

The way we actually worked was not me asking for snippets. I'd describe the family behavior I wanted, and Codex would go trace what already existed across the controller, the endpoint service, the browser extension, and the packaging scripts, then come back with what had to change in each one. That mattered more than code generation did, because this is a multi-process product. Something that sounds trivial like "pause the internet" touches the policy engine, the cached rules sitting on each device, the parent dashboard, and the extension's decision path. Keeping those four in agreement by hand is where I would have lost weeks.

What we built together: a controller in Python that keeps family policy and activity in a local SQLite database, with no vendor account and no hosted control plane anywhere in the path. Devices enroll once using a high entropy code that expires in fifteen minutes and cannot be reused. After that, every endpoint check-in is signed with per-device secrets and validated against replay using nonces, and prompt payloads are encrypted per device with Fernet before they ever cross the LAN, because families run plain HTTP inside their own network and that text is still sensitive. On Windows, retained evidence is protected with an OS key. LAN discovery announces that a controller exists and nothing more. It cannot enroll anything on its own, which was a deliberate decision after Codex and I talked through what an attacker on the same coffee shop Wi-Fi could do.

The browser extension is Manifest V3 and only ever talks to a loopback bridge on the same machine. It evaluates navigation, form text, chatbot prompts, text file uploads, and visible page time. If the controller goes quiet for a bit, the cached domain, schedule, pause, SafeSearch, and prompt rules keep making decisions locally, so protection does not evaporate the moment my controller reboots.

For the GPT-5.6 Safety Plan Studio, the design argument came first and the code came second. I wanted AI help translating "calmer school nights and a fair way for him to ask" into actual categories and schedules, but I did not want a model with its hand on the enforcement switch, and I did not want my kid's browsing history feeding a prompt. So we split it: a preview route that builds a minimized request (age band, my goals after local secret and contact redaction, the current rule settings, and the category names available on my controller) and a completely separate apply route that only a parent can trigger. Requests use strict Structured Outputs with store set to false and a privacy preserving safety identifier. Child profile fields, browsing history, prompt content, device activity, and the family database are never attached. We later added a second way in for parents without an API key, where the request runs through the parent's own Codex CLI sign in so the ChatGPT OAuth stays inside OpenAI's tool and my app never handles a credential at all.

The QA is the part I did not expect. During live browser testing Codex caught two things my own happy path clicking had sailed right past. The first was concurrent dashboard reads sharing an unlocked SQLite connection. The second was subtler and worse: a category catalog mismatch that would have quietly dropped an existing drugs block when a parent applied an AI generated plan. That is exactly the kind of bug that erodes trust in a safety product, since the parent thinks they tightened things and actually loosened them. Both have regression tests now.

We also did the unglamorous work. Self contained Windows executables for the controller and endpoint so nobody needs Python installed, Windows service wrappers, a macOS endpoint DMG build script, a judge demo mode that seeds clearly labeled synthetic activity while still using the real database and policy and apply paths, and documentation that states plainly what this cannot enforce yet. The browser extension is the current enforcement surface. OS wide filtering is not done, so I say so instead of implying coverage I don't have. Signing and notarization are still ahead of me and that's written down too.

The thing I took away from building this way is that AI was genuinely excellent at planning, tracing, and reviewing, and I would not want it holding authority over what my kids can see. That belief ended up shaping the whole architecture, not just the AI feature.

Robert Greenwood
A single father, freelance programmer with a total of 18 years of experience.

Updates

Robert Greenwood started this project — Jul 21, 2026 11:08 AM EDT

Leave feedback in the comments!

Log in or sign up for Devpost to join the conversation.

Built With

Updates

Submission history