If you’ve been following the HYNAWEB journey, you know we build enterprise-grade cloud infrastructure entirely on a $150 Android phone using Termux.
But building on mobile isn't just about writing code; it's about battling the "Termux Poltergeist." Yesterday, a corrupted string literal in my terminal accidentally broke the @hynaweb/kavlr npm package. After debugging the source, republishing v1.0.9, and accidentally letting my own AI security gate block me from deleting my demo file, I finally recorded flawless, bug-free terminal demos for both KAVLR and HYNA-Stage.
No shaky phone videos. No heavy video editors. Just raw, unscripted Asciinema terminal recordings.
Here is what they actually look like in action.
🛡️ 1. KAVLR v1.0.9: The AI Action Gate
KAVLR is a fail-closed terminal leash for autonomous AI agents. It uses a 90-article Constitutional AI framework to grade the action of a command, not just the text.
Below is the flawless, bug-free recording of the CLI booting up, rendering the ASCII art, and verifying the security status.
(Note: During the recording session, KAVLR actually intercepted and blocked me from running rm kavlr-demo.cast to delete the file, giving it a 0/100 score. The Self-Preservation Protocol is real!)
What you're seeing:
-
kavlr --help: The clean CLI routing. -
kavlr init: The beautiful cyan ASCII art rendering. -
kavlr status: The greenSECURE (ACTIVE)andONLINEtelemetry.
☁️ 2. HYNA-Stage: 5.1s Ephemeral Cloud Sandboxes
While KAVLR protects the terminal, HYNA-Stage gives AI agents a safe place to execute code. It provisions isolated Firecracker microVMs using a pre-warmed pool to eliminate cold starts.
Here is the live recording of the 5.1s warm-start benchmark:
What you're seeing:
-
hyna-stage login: Secure JWT authentication. -
hyna-stage up index.html: The CLI hitting the Supabase Edge Function, provisioning the E2B sandbox, and returning a live preview URL in exactly 5.1 seconds.
🧠 The Reality of Building on a $150 Phone
Most developer tools are built on $3,000 MacBooks with perfect internet connections. Building HYNA-Stage and KAVLR on a budget Android phone means dealing with paste-buffer corruptions, strict API gateway blocks, and mobile terminal quirks.
But it also forces you to build lean, fast, and highly optimized tools.
The Termux Poltergeist is defeated. The npm packages are clean. The demos are live.
Try them yourself:
-
KAVLR:
npm install -g @hynaweb/kavlr -
HYNA-Stage:
npm install -g @hynaweb/stage - Hacker Stage Jam: hacker stage
What should we build next? Let me know in the comments! 🐯☁️🚀
Top comments (10)
Hi CEO Harun, I think you are doing a great job. But I have a few tips.
Tips:
ForkCount
CEO of forkcount.github.io
Hey nicky
I really appreciate this. You just hit on the single biggest challenge I face right now: The Maker's Trap.
You are 100% right. I’ve been so hyper-focused on shipping new architecture (like the 90-article AI constitution and 5.1s warm pools) that I’ve neglected the central portfolio and the distribution side of things. Five products with massive, active traffic are infinitely more valuable than ten products sitting in the dark.
Since you’re already analyzing my build cadence, thinking strategically about traffic, and calling yourself the CEO of Forkcount... I have a question for you:
How would you feel about officially collaborating on the growth and distribution side of HYNAWEB?
We have the technical engine (KAVLR, HYNA-Stage, Hacker Stage Jam), but we need a dedicated growth mindset to get these tools the audience they deserve. I’d love to pick your brain on setting up that central portfolio website and a real distribution strategy.
Let me know if you’re open to a chat. We could use a CMO. 🐯☁️🚀
bro what do you think 🔥 @adamthedeveloper , @effessdev , @forkcount , @jsb-securedme ,
But what is KAVLR?
Great question, EffessDev! 🐯
KAVLR is a fail-closed, AI-powered terminal leash designed specifically for autonomous coding agents (like Cursor, Aider, or Claude Code).
The core problem: AI agents sometimes hallucinate and try to run destructive commands (like
rm -rf, modifying.envfiles, or exfiltrating data viacurl). Traditional text-based guards are easy to bypass with obfuscation (likebase64).How KAVLR fixes this:
Instead of scanning the text of a command, KAVLR intercepts the execution and grades the semantic action using a 90-article Constitutional AI framework.
If the AI tries to do something dangerous, KAVLR gives it a
0/100score and blocks the execution entirely before it ever hits the shell. (That’s what you saw in the demo when it blocked me from deleting my own file! 😅)It’s basically a bouncer for your terminal, ensuring your AI agent can build fast without accidentally destroying your codebase.
Want to test it on your own machine?
👉
npm install -g @hynaweb/kavlr👉 Repo: github.com/harun-sket/kavlr-api
Would love to hear what you think of the architecture! 🛡️☁️
this is not what AI hallucination means, by the way.
You're right on the strict terminology—"agent misalignment" or "unintended destructive execution" is more accurate than "hallucination" in this specific context. Thanks for the catch.
The core engineering challenge remains the same, though: whether it's a statistical glitch or a bad prompt chain, the shell needs a deterministic, fail-closed guardrail to prevent catastrophic actions. That is exactly what KAVLR provides at the execution boundary.
Appreciate you reading the thread and keeping the technical standards high! 🛡️
oh hi, mr CEO
Why do you call me mr CEO bro ?
I'm 12 and built this entirely on a $150 Android phone using Termux. The hardest part was handling the Supabase API Gateway blocking my custom headers, and getting the MicroVM to boot a web server before the E2B URL loads. For the senior devs here: what would you add to the architecture to make it more reliable for production?