Summary

• Infrastructure Realignment: Transitioned from stateless Cloudflare Workers to a dedicated VPS on Hetzner, moving away from "flawed" containerization to allow the agent full bare-metal autonomy and stability.
• Security Hardening: Adopted a "Security Always" posture by closing all ports and utilizing Cloudflare Tunnels to enforce Zero Trust authentication, effectively shielding the agent behind a private login layer.
• Stability Optimization: Resolved persistent deadlocks and compaction crashes by fine-tuning technical configurations, including increased token limits and extended timeout thresholds for complex processing.

10x productivity...? Not now.

Oh the irony. Since launching Pinchy, I thought that my productivity was going to increase tenfold. Instead, I spent many sleepless nights just attempting to fix it. Rather than gatekeeping the information, I'm sharing my journey to help you better develop your OpenClaw assistant.

What was learnt from CloudFlare?

Stateful product on stateless infra. Dumb.

Cloudflare's solution was a good introduction towards a safe deployment of OpenClaw. However, it lacked reliability. Using stateless servers to achieve stateful manner is counter productive. Pinchy either didn't run at scheduled time, kept crashing, or killed itself through misconfiguration. The most powerful aspect of this bot wasn't working, turning it into another chatbot.

Fixed. Oh. Containerisation is flawed.

To counteract this issue, I decided to start anew with a stateful server like VPS on hetzner and run on container. But even then, the same issues occured. I realised... the container implementation of openclaw is flawed. Openclaw was intended to run on baremetal with full autonomy. I was holding back the potential of Pinchy due to fears on security.

Security first? Security always.

While thinking of how to solve the issue of security, I recalled CloudFlare's implementation for OpenClaw and adopted best practices such as closing off all ports and exposing the endpoint via CloudFlare tunnel. This allowed for zero trust authentication, strengthening security for my server. If you tried to access Pinchy on my website, you would be blocked by a login page.

Following OpenClaw best practices, rather than exposing endpoints for webhooks, I implement polling instead. Although the same security risks still exists such as being prompt injected, all I can do now is to add more layers of security, while maintaining the same level of functionality.

Crash again? Oh, it was just configurations.

Despite doing all these, Pinchy still kept crashing due to compaction and deadlocks. I increased the token limit, increased timeout for deadlocks, and increase reserved tokens for compaction.

Finally, I can rest.

Finally, Pinchy is stable and able to start doing more complex tasks.

Conclusion

Real 10x gains? We will see...

After getting Pinchy stabilised, I've been adding more and more useful tools to help it accelerate the chores of my life. These are my top three skills now that I use.

Homework Assistant

Pinchy reads uploaded files on task, researches online, draft in Google Document, sends it to Sapling AI to humanise, then reformat the Google Document. Good for homework chores and drafting.

Contrarian Sniper

Pinchy looks online daily for contrary sentiments, using deep value analysis and risk assessment to give a conviction score. It's designed to be conservative with constrained portfolios, not for speculation.

Senior Software Engineer

Coding

Pinchy is able to act like a senior software engineer, understanding the flow of the code from end to end. It leverages existing documentation, preferring reusability and simplicity over writing code. Most of its time is spent planning, and finding ways to reuse code.

Reviews

Pinchy reuses an open source PR review skill and adds a twist of its own, tracing from end to end to give better code reviews and suggestions.