OpenClaw · self-hosted AI agent

Everything my OpenClaw does for me

Vatsal Mishra · Applied AI @ Google

meet Cloidberg, my always-on agent

Chief of staff Builder Comms operator Home OS Media pipeline

How I got here

I tried to build this before I knew what the missing piece was.

🧪

The first attempt

When local LLMs and Ollama got good, I wired Telegram to local inference. It replied, but it did not remember, recover, or act.

🦀

The click

Then I saw Peter launch OpenClaw on X and thought: this is exactly what I was trying to build.

Mini PC Telegram Discord Gmail WhatsApp RO Calendar RO Home OS Browser Local LLMs

What it does

One agent. Many hats.

Three roles that matter most

🧭

Chief of staff

Morning briefing, project management and context, reminders, and memory.

🤝

Personal assistant

Email management, WhatsApp checks, Home OS, receipt-to-pantry, recipes, and household admin.

🛠️

Builder assistant

Reads code, patches files, runs builds, and delegates to sub-agents and AI coding tools.

And beyond those, it also

📸

Idea to reality

Spins up quick MVPs, products, even deep research and analysis, handed back in a digestible way. Just by talking to it from my phone.

🎬

Content machine

FakeOut blog drafts, Field Notes, newsletter digest, Remotion videos, captions, reels, and social drafts.

🛰️

Personal infrastructure

Runs my homelab: Home OS, Mission Control, media library, self-hosted cloud storage, isolated workspaces, and guardrails.

Real work, every day

Across every surface I actually use.

Generating a LinkedIn asset from Discord — Social Media Assets

SEO-optimized blog posts written for FakeOut after approval — SEO blog posts

Driven from a clean UI when I want control, or a single Telegram message when I want speed.

Under the hood

How I keep it lean: tokens, memory, cost.

Same machine, same agent, but disciplined about what enters the context window and which model does the work.

Right-size the model Main chat on a capable model; digests, scouts, and heartbeats routed to cheaper, faster ones. The heavy model is reserved for heavy lifting.

Delegate to sub-agents Heavy builds and research run in isolated sub-agents, so the main chat stays short. Short sessions hit the prompt cache; long ones burn it.

Tiered memory Raw daily notes distilled into a small always-loaded memory, the rest archived and pulled on demand. Context stays lean every session.

Context just-in-time Tool and project docs are indexes; detail is read only when a task needs it. Separate workspaces stop context from bleeding together.

What is possible

One person. One Mini PC. A resident agent that actually helps.

I used to sit on a backlog of ideas and ship almost none. Now I spin up a prototype in minutes and find out if it is worth pursuing.

Scan → deck, socials, blog, and FakeOut

Click here for all the links

Take the links with you

Huge thanks to Give(a)Go for running the OpenClaw meetups and growing the community here in Dublin.