KingxTech: what we shipped this month

Nov 15, 2026 · @Alkhassim Lawal Umar

The short version

In the last month, KingxTech grew from a site builder with an AI chat into a platform that runs backends, guards its own AI spend, and plugs into Claude, ChatGPT and Gemini.

KingxTech is a platform I'm building in Kaduna, Nigeria, for creating, hosting and publishing web apps with an AI that works inside your project. It's a React and Vite frontend, Supabase for auth and data, and an Express API on Vercel. I'm the only developer, so this post is half changelog and half notes to myself about what I'd do again.

Since October 15 that came to about 160 commits on the platform branch. Here is what shipped, grouped by what it means for someone using the platform.

Backends, hosting and the workspace

The biggest change is that a project can now have a server, not just a frontend.

• Project servers (Pro). A Node API that sleeps when idle and wakes on the first request. Project code runs as a restricted user behind a signed gate inside the sandbox, with usage accounting and plan limits on top.
• Project Data API. Published sites can save and read their own records, so a static site no longer needs a separate backend for simple data.
• AI for your site. A public chatbot endpoint, so a published site can offer AI without ever holding an API key.
• Restore points. An automatic snapshot after each publish, plus manual save, restore and delete.
• Storage limits and a usage page. Free 25 MB, Pro 250 MB and Pro+ 1 GB, enforced on every file write, with a per-project meter.
• Site visits. Privacy-friendly daily page views for published sites.
• Starter templates. Landing page, portfolio, link page, shop and dashboard.
• K-XpertAI beside your code. The assistant docks next to the page and shows files being written live.
• A docs section. Search, video embeds and ten guides, plus an API guide so the AI assistants know the platform too.

Two product updates went out on the newsroom around the same time: backend support and NeuroCore, private AI that works offline.

Phase 4: keeping AI affordable and safe

Phase 4 was about making sure no single user, and no single stuck request, can drain the AI budget.

• Usage caps are enforced. AI usage limits now run in enforce mode, so the plan limits are real rather than advisory.
• A concurrency guard. Each user can only have so many paid AI runs going at once.
• Billing for runs that die early. If a run crashes before it records its cost, it is still billed, so a failure can't become a free request.
• A server usage card. Users see how much of their server allowance they have used, visitors see a friendly paused page when a server is out of allowance, and admins can adjust an allowance.
• A real Model Router. Free plans default to smaller open models, paid plans reach larger ones, and every route has spending guards.

Security got its own pass at the end of October: the terminal tool, the SQL console, rate limiting, dependencies and response headers were all hardened. On the trust and safety side, there is now an admin dashboard with a support inbox, account suspension, and a flags queue that reviews AI misuse, plus a Terms of Service that every account accepts.

Custom MCP: Claude, ChatGPT and Gemini can work in your projects

KingxTech now runs an MCP server, so an AI app that supports custom connectors can sign in to your account and work with your projects directly.

You paste https://site.kingxtech.name.ng/mcp into the app, sign in to KingxTech, and approve it once. The sign-in is Supabase's OAuth server, with a KingxTech consent page in front of it. After that the app gets six tools: list projects, get project status, list files, read a file, write a file, and check plan and storage usage.

The boundaries matter more than the tools, so they are shown on the consent page and in Account > Connected apps:

• It can: see your project list and each project's publish and server status, read text files (secrets files like .env are hidden), create or replace text files in projects you can edit, and see your plan and storage use.
• It cannot: delete files or projects, publish sites or start servers, see or change environment variables and secrets, or touch billing, domains or account settings.

Tokens issued to these apps are scoped, so they can reach the connector endpoint and nothing else. Removing an app in Connected apps ends its access.

I've connected it to Claude, to ChatGPT on a free account, and to Gemini. In ChatGPT, custom connectors are added from the web or desktop app rather than the mobile app, and plan availability can change. Once connected, it works from the phone apps too.

One bug worth admitting: the Connected apps list sat on loading skeletons forever. I was calling oauth.getUserGrants(), but supabase-js names that method listGrants(). The call threw, and my page had no failure state, so it looked like it was still loading. The fix was one word plus a try/catch.

Community and partnerships

The platform only matters if people can learn to use it, so part of the month went to teaching.

KingxTech is partnering with Kitoko Hub on their AI & Tech Creators Workshop, starting in February 2027 with 50 seats. Learners work inside the KingxTech ecosystem, and KingxTech provides technical support, tools and mentorship. The application page for it is itself a project hosted on KingxTech, which is a good test of the hosting. The announcement is on the newsroom.

There is also a growing docs section with video guides, and a Discord for questions and feedback.

What I learned building this solo

1. Put spending guards in before the AI is public. Caps, a concurrency limit and billing for crashed runs are boring work, but they are what lets a small team offer AI at all.
2. Treat an AI app like a stranger with a key. For the MCP connector I started from what it must never do (delete, publish, read secrets, touch billing) and only then listed what it can do.
3. Every loading state needs a failure state. My Connected apps bug was invisible because a thrown error looked exactly like a slow request.
4. Check scheduled jobs against the database, not the dashboard. The daily plan-expiry cron on Vercel's Hobby plan runs somewhere inside its scheduled hour. I confirmed it worked by looking at which accounts had dropped to the free plan and when, not by trusting the schedule.
5. Small review loops beat big plans. Most of the month was a pull request, a test on my phone, and the next pull request.

What's next, and how to try it

Next I want KingxTech listed in the connector directories of AI apps, so people can find it without pasting a URL. That needs a privacy policy page, a support contact and a reviewer demo account. I'm also settling plan limits and a cap on how many project servers can run at once across the platform.

You can try everything above with a free account at kingxtech.name.ng. The docs cover the connector setup, and the Discord is the fastest way to tell me what breaks.