In April I wrote about OpenClaw — my self-hosted AI agent on a Hostinger VPS, wired to Telegram, email, calendar, and GitHub. That post ended with a short "What's Next" list. Three months later, most of it shipped — and the setup feels less like a chatbot and more like a small ops team.
What's new since April:
- Blog posts — voice note in, researched draft out, published on my site
- Code from GitHub issues — agent plans, I approve, PR lands
- Invoices — bilingual PDFs for US and German clients, email draft ready to send
- SEO & compliance work — live data pulls and reports without me clicking through five tools
- One Telegram chat, many projects — separate topics so context doesn't mix
- Conference proposals — talk drafts from memory, not from a blank page
Below: what changed, what broke, and what I'd do again.
Quick recap
OpenClaw is a persistent assistant with memory. I talk to it on Telegram — often by voice. It can run tasks, read my repos, draft email, and check GitHub. It only sends mail after I say yes.
Everything below builds on that.
1. Blog posts — from idea to publish
The April post promised blog drafting. That happened — repeatedly:
| Post | What the agent did |
|---|---|
| Hetzner EU production series | Researched the stack, drafted EN+DE, coordinated images |
| Lovable → Hetzner DACH | Same pattern, tied to a real client story |
| SISTRIX MCP SEO audit | Pulled live SEO data, read the client's site code, exported a strategy PDF |
Typical flow: voice note → agent gathers context from past notes and repos → markdown draft → I edit for tone and accuracy → publish.
Lesson: Great at first drafts with real data. My job is "would I post this under my name?"
2. GitHub issues → plans → PRs
The biggest upgrade: OpenClaw can hand off coding work to Cursor in the actual repo.
Today:
- New issue on a client's product repo
- I get a Telegram ping
- Agent writes an implementation plan from the codebase
- Plan lands as a GitHub comment — visible to the team, not buried in chat
- I approve → code gets written → PR opens
I stopped burning the expensive model on every small check. Quick triage stays cheap; planning and coding use the heavy models only when it matters.
Lesson: Keep planning in the repo and in GitHub. Telegram is for pings and decisions, not for hiding the plan.
3. Invoices — the boring work I avoided
Invoice day used to mean copy-paste HTML, hunt for rates, second-guess tax wording, export PDF, upload, draft the cover email. Now I open the right Telegram topic and say prepare the next invoice.
The agent already knows each client's rules — US vs German format, currency, VAT, layout. It pulls the last invoice as a template, bumps the number, fills in the period and line items, and asks me only what's missing. Output: PDF in Dropbox, email draft waiting — send only after I approve.
Lesson: Highest ROI automation I didn't mention in April. Repetitive, rule-heavy, easy to get wrong. Perfect for an agent with memory.
4. SEO, WordPress, compliance — one prompt, one artifact
April: basic memory and a few tools. July: the agent can talk to external services I use every week — SEO data, a client's WordPress blog, security scan reports.
The pattern is always the same: one Telegram message, agent combines live data + code context, I get a dated deliverable (markdown, PDF, or issue comment). I wrote up the SEO workflow here.
Lesson: One integration per job you do manually more than once a month. Don't wire everything on day one.
5. Telegram topics = project switcher
One group chat, many topics — blog, US client, German client, side projects.
Each topic loads the right context: which repo, which tone, which rules. I don't mix invoice talk with blog drafting in one thread.
Lesson: If your bot goes quiet in a group, check you're on a recent OpenClaw build. Forum topics had a reply bug that's fixed now.
6. Conference talk drafts
Between client work I still submit to AWS Community Days, AgentCon, CodeMotion, re:Invent. The agent keeps proposal drafts in a folder and knows the submission format.
Typical flow: draft a talk on security scanning for Poland Community Day → agent pulls from past project work and prior talks → title, abstract, outline ready to paste into the submission form → calendar reminder for the deadline.
Lesson: CFP writing is research + structure + your voice. The agent does the first two; I keep the third.
Numbers (rough, three months in)
- GitHub: dozens of issues triaged; several fixes shipped via agent → Cursor → PR
- Blog: 4+ substantial posts with agent help (EN+DE pairs)
- Invoices: 5+ US sprint invoices + German client invoices — PDF + email draft each time
- CFP: 10+ proposal drafts across Poland, DACH, Milan, re:Invent
- Cost: picking the right model for the task matters more than VPS hosting
- Failures: one email sent too early → hard "ask before send" rule; morning briefing broke once after an update — fixed with a version bump and a cheaper scheduled job
What's next
- Auto-implement after explicit GitHub approval
- Infra changes from chat (Hetzner work in progress)
- Quarterly VAT prep from the invoice folder — agent-assisted, accountant still reviews
Conclusion
Three months ago OpenClaw was email, calendar, and GitHub pings. Now it's an orchestrator: topics for context, Cursor for code, live integrations for data, the blog for content, invoices and talk drafts for the business side.
Still on €10/mo Hostinger Docker.
If you already run OpenClaw, what helped me most: hand off coding to Cursor, one Telegram topic per client/project, and one external integration per repetitive manual job.
Links:
Questions or want help wiring your agent? office@martinmueller.dev or calendly.com/martinmueller_dev.
Course: OpenClaw as a personal AI OS
This post is what shipped after three months. The course is the build — a personal AI on your VPS, from gateway install through the GitHub PR loop.
You install the OpenClaw gateway, wire Telegram and Slack, and keep the agent continuous with SOUL.md, MEMORY.md, and daily notes. Heartbeats and cron for proactive checks. A Cursor agent loop that opens GitHub PRs. MCP servers for calendar, email, and search.
OpenClaw: Personal AI OS — two-day team training, on-site or remote, price on request.
