OpenClaw in Stuttgart: a builder's guide to running your own agent

We ran the OpenClaw Hands-On Stuttgart workshop with about fifty builders, one shared goal, and a single rule: by the end of the night, your laptop is talking to an agent that lives in markdown files you can read. This post is the short version of the talk, plus the path you can follow on your own. All slides, workflow starters, and deployment recipes live in the openclaw-guides repo.

What is OpenClaw?

OpenClaw is an open-source agent gateway. It routes messages from any channel (web UI, CLI, Telegram, Discord, Slack, WhatsApp) to an AI agent and back, and persists long-term memory as plain markdown files on disk. It's local-first, MIT-licensed, and ships with the Pi coding agent (badlogic/pi-mono) embedded, so the assistant can read, write, edit, and run shell commands.

• Gateway — one normalized hub between humans and agents.
• Markdown memory — SOUL.md, USER.md, MEMORY.md, notes/. You can cat, vim, and git diff your agent's mind.
• Pi inside — Read, Write, Edit, Bash. The agent does things, not just chats.

OpenClaw vs Claude Code, Cursor, Hermes, and Pi

These names get mixed up constantly, so here's the short answer.

• Claude Code and Cursor are coding assistants you use inside an IDE or terminal. They don't run cron jobs and they don't expose themselves over Telegram.
• Pi (badlogic/pi-mono) is a coding agent toolkit by Mario Zechner. OpenClaw embeds Pi as the agent runtime; Pi by itself is closer to Claude Code than to OpenClaw.
• Hermes by Nous Research is the closest peer to OpenClaw: an open-source, self-hosted agent with persistent memory, a multi-platform gateway, and 40+ built-in skills. Python-based, very active in 2026.
• OpenClaw is the gateway plus markdown memory plus Pi, in TypeScript. If you want a self-hosted Claude Code alternative that runs always-on and reaches your phone, this is the shape.

What we covered in the Stuttgart Hands-On

• The five concepts: Gateway, Channels, Agents, Sessions, Memory.
• Three workflow tracks: a news briefing agent (beginner), a price monitor (intermediate), and an email monitor (advanced).
• Where to run it: laptop, coral.inc, or a VPS. We picked one per table.
• Live demo: message in, markdown file written, reply out.

The slides, the facilitator script, the three starter projects, and the full troubleshooting guide are in sachin-aag/openclaw-guides. Clone it once and you have everything from the room.

Your first workflow in 10 minutes

The hello-world isn't a chatbot. It's a news briefing agent that fetches RSS feeds, builds a structured digest, and writes it to a markdown file you can edit.

git clone https://github.com/sachin-aag/openclaw-guides.git
cd openclaw-guides/workflows/news-briefing-agent
npm install
cp .env.example .env   # add your model API key
npm run briefing

Add a model API key (Anthropic, OpenAI, Google, or Featherless), run the script, and a new file shows up under workspace/briefings/. That's your agent's output, in a file your editor can open.

Is OpenClaw a Claude Code alternative?

Sort of. Claude Code is one assistant in your terminal. OpenClaw is the gateway around an assistant — same tool primitives (Read, Write, Edit, Bash) via Pi, but reachable from chat apps, runnable on a cron, and with memory you can grep. If you want Claude Code that lives on a VPS and messages you on Telegram, that's the OpenClaw shape.

Where to run it: local vs coral.inc vs VPS

Three realistic targets. Most people start local, prove the workflow they care about, then move to a VPS once cron actually matters.

Local laptop

Free, fastest to iterate, perfect for editing memory files in your IDE. Big caveat: laptops sleep, so cron only fires when you're awake. Treat it as a dev environment, not a deployment.

coral.inc

Managed host with a five-minute setup. About $50/month base; the workshop code STUTTGART knocks $30 off the first month. Pick this when you want to demo to non-technical teammates without explaining what a VPS is.

VPS (Hostinger or GCP Compute Engine)

Around €5-15/month, takes 30-60 minutes the first time. This is the right home for an always-on agent: cron actually fires, your memory persists, and you control the firewall. Full deploy guides for Hostinger and GCP are in the repo.

Memory as markdown changes how you build with agents

Most agent frameworks make memory a black box: a vector index you query, a hidden conversation log, a managed graph. You can't read it. You can't grep it. You can't git diff it.

OpenClaw goes the other way. Your agent's mind is a folder.

• SOUL.md — personality, voice, operating principles. Open it, change a sentence, save. The agent thinks differently next turn.
• USER.md — what the agent knows about the human(s) it talks to.
• MEMORY.md — long-term notes the agent appends over time. Decisions, recurring patterns, anything tagged #keep.
• notes/ — daily logs and scratchpads, anything domain-specific.

Three things become possible that vector-DB agents struggle with: you can audit the mind out loud at a meeting, you can edit it without retraining, and you can version it in git and roll back when the agent develops a weird habit.

Pi as the coding brain

OpenClaw doesn't reinvent the agent loop. It embeds Pi (badlogic/pi-mono) directly. Pi gives the assistant four primitive tools:

• Read(path) — open files, scan folders.
• Write(path, content) — create new files.
• Edit(path, patch) — surgical patches to existing files.
• Bash(cmd) — run anything in a shell. This is powerful and dangerous; treat your tool policy like a security boundary.

That's why an OpenClaw agent can write to MEMORY.md, run a Python script, check git status, or compose an email without you wiring custom tooling. Functionally, it's a junior dev with shell access and a markdown notebook.

Common pitfalls we hit at the workshop

• Cannot find module … — run npm install inside the workflow folder, not the repo root.
• Script runs but nothing gets written — check the output directory in .env exists and is writable.
• 401s presented as 500s — most providers swallow auth errors. Verify the key matches OPENCLAW_PROVIDER.
• Cron doesn't fire on my laptop — laptops sleep. Use caffeinate -i on macOS or move to a VPS.

The full list lives in the openclaw-guides howto folder.

Where to go next

• Add a second channel — wire Telegram or Discord so you can talk to your agent from your phone.
• Multi-agent routing — run a researcher and a writer that hand work to each other.
• Hybrid model routing — cheap model for routing, expensive one for thinking.
• Package one workflow as a reusable Skill so other agents can invoke it.

If you're in Stuttgart, the next CREATORS sessions are listed on the events page, and more practical write-ups land on the blog.

FAQ

What is the Pi coding agent?

Pi is an open-source coding-agent toolkit by Mario Zechner, distributed as a TypeScript monorepo at badlogic/pi-mono. It includes a CLI coding agent, a unified multi-provider LLM API, an agent runtime with tool calling, and a terminal UI. OpenClaw embeds Pi to handle the actual agent loop.

What's the difference between OpenClaw and Hermes?

Both are open-source, self-hosted agent runtimes with persistent memory and a multi-platform gateway. Hermes (by Nous Research, released February 2026) is Python, has 40+ built-in skills, supports automated skill creation and multiple execution sandboxes (local, Docker, SSH, Modal, Singularity). OpenClaw is TypeScript, ships Pi as the embedded coding agent, and stores memory as plain markdown you can git diff. Pick OpenClaw when you want minimal moving parts and editable memory; pick Hermes when you want a richer skills catalog and stronger sandboxing out of the box.

Can I self-host OpenClaw without a VPS?

Yes, on a laptop or a Raspberry Pi-class home server. You only need a VPS when always-on cron, real users, or external webhooks matter. The runtime itself is small.

Which open-source AI agent should I pick in 2026?

If you want to build coding workflows on your own files: Pi or Claude Code. If you want a self-hosted assistant reachable from chat apps with memory you control: OpenClaw or Hermes. If you want a managed framework inside a Next.js app: the Vercel AI SDK plus Workflow DevKit. They solve different shapes of problem; mixing two of them on one stack is normal.

Does OpenClaw work with Telegram, Discord, Slack, and WhatsApp?

Yes. Channels are pluggable. The workshop covers the web UI by default and Telegram/Discord as the easiest second channel. Slack and WhatsApp need more setup (bot tokens, webhook verification) but follow the same pattern.

How is markdown memory different from a vector database?

A vector DB stores embeddings you can semantically search but not read. Markdown memory is just files: you read them on each turn, you edit them by hand, and you version them in git. You give up some retrieval cleverness; you get auditability, edit-ability, and zero infra.

Is OpenClaw production-ready?

For personal use and small teams, yes; many builders run it on VPSes today. For high-traffic public products, treat it like any agent system: add rate limiting, observability, secret rotation, and a tool policy that doesn't include unrestricted Bash.

Further reading

Last updated: May 2026.

• Recap of the first OpenClaw meetup in Stuttgart (March 2026)
• sachin-aag/openclaw-guides — slides, workflow starters, deploy guides
• openclaw.cc — official site
• docs.openclaw.ai — official docs
• badlogic/pi-mono — the Pi coding agent monorepo
• hermes-agent.org — Hermes by Nous Research
• OpenClaw Hands-On Stuttgart on Luma