What we are building today

A self-hosted agent server you own and control. By the end of this talk you will know how to build one from scratch. Everything you need is linked below.

You own the config, you own the exit

What that script would find:

All plain text or SQLite. No proprietary formats. No vendor lock.

What this talk is and isn't

Tailscale

A private mesh network across all your devices. Runs alongside everything you already use without conflict.

• Works on iPad, iPhone, Windows, Mac, and Linux with zero configuration changes to your existing network
• Every device gets a stable private IP that works regardless of where you are
• All your hosts stay accessible to each other but completely off the public internet
• Built on WireGuard: fast, lightweight, and always on in the background
• Advanced features like exit nodes let you route traffic through a specific device when you need it

Getting started

Seven steps from zero to a working agent server. Details in the deployment guide.

1. 1
   Set up a Tailscale account

   Free at tailscale.com. This becomes your private network.

2. 2
   Get an Anthropic API key

   Sign up at console.anthropic.com and generate an API key.

3. 3
   Build or rent a server

   Mac Mini at home, or a VPS from Hostinger ($10-30/mo). Select Ubuntu.

4. 4
   Log in as root and install Tailscale

   curl -fsSL https://tailscale.com/install.sh | sh
   sudo tailscale up

5. 5
   Install Claude Code

   npm install -g @anthropic-ai/claude-code

6. 6
   Log in to your Claude account

   claude login

7. 7
   Install Claude Workspaces management

   git clone https://github.com/mikeengleai/openclaw-framework.git

The order to build in

This is the roadmap for the rest of the talk. Each step builds on the last.

1. SSH + Tailscale (secure access to your box)
2. tmux + Claude Code + cw (persistent development sessions)
3. One agent that solves one painful thing
4. Slack binding (give it a channel)
5. Cron a daily report (make it proactive)
6. Add a second agent (now you have a pattern)

Four ingredients

Claude Code Agent View

Press the left arrow in Claude Code to see all sessions in one terminal window. The terminal-native counterpart to the web-based /remote-control.

Both have their place. Web is more flexible for mobile. Terminal is faster at the desk.

Three lessons

Channel bindings

The daily system map

Every day at 5:30 AM, a cron job regenerates a full markdown snapshot of the entire system: agents, skills, bindings, cron jobs, memory files, Tailscale nodes, security config.

20 sections. Covers infrastructure, agents, skills, cron, memory, security, plugins, Tailscale, scripts. The system map for this talk was generated on 2026-05-23.

What QMD is

SQLite-backed per-agent memory with structured frontmatter, queryable across agents.

Proof it's in use

Real file sizes from production:

9 memory files total on production. The youtube_kb database holds months of transcripts from subscribed channels. scout holds competitor intelligence. These are working databases, not demos.

AI Security Framework

Keeping tabs on what competitors and partners are doing in the AI security space. The agent scours sources daily so you don't have to.

Sources scoured daily

150+ entries across 127 stories from 11+ domains on production. Source count grows automatically.

Meeting intelligence for the whole company

Meeting transcripts locked inside Otter's app are invisible to anyone not in the room. This agent makes them searchable from Slack.

Email, calendar, and HubSpot in one place

Data fragmented across three systems. The EA spent half her time switching tabs.

Peptide research database

The browser problem

Self-hosting means you decide what the agent can access. The responsibility is yours.

The agent-browser skill

• Persistent browser context per agent
• Cookie store on disk (survives restarts)
• Screenshot artifacts for every navigation
• Xvfb virtual display on the server
• 10 scripts in the linkedin-feed skill alone

~/.openclaw/skills/agent-browser/cookies/
├── linkedin-engle/
├── mybrowser/
└── youtube-kb/

The LinkedIn workflow

• Cookie cycling to maintain authenticated sessions
• Daily feed monitor (7 AM and 7 PM UTC cron)
• Company page monitoring and engagement tracking
• Post draft pipeline with human review

YouTube knowledge base

The apartment dashboards

Boston and Pennsylvania rental markets. Same pattern: scrape structured data, store in SQLite, serve a dashboard.

The publish-static skill deploys any directory of HTML/CSS/JS to a Tailscale-accessible URL.

Honest takes

Cost reality

The Daily Model Health Report cron tracks API usage across providers. Cost varies with workload: a heavy week of YouTube ingestion or AISF sweeps pushes the number up. Quiet weeks drop below $100.

Per-agent model choice as defense in depth

• Email and calendar (Opus 4.7): highest trust tier, US-based provider, strongest safety guardrails
• Security analysis (Sonnet 4.6): same trust tier, lower cost for the volume
• Content curation (Gemini Flash): non-sensitive, high volume, cost-optimized
• Transcript processing (DeepSeek): no customer data, no strategic content, pure text transformation

The canary pattern

Architectures kept similar on purpose (same Slack patterns, same skill structure) so issues surface on canary before they reach production.

Always-on monitoring

Honest position on backups: no restore drills. Host-level backups taken through the Hostinger interface before major upgrades.

Things to do differently if starting over

What's still fragile

• Browser cookie expiration is a single point of failure for authenticated browsing workflows. When cookies expire, the agent loses access and needs fresh credentials manually.
• No formal eval suite per agent. A quality-sentinel skill exists on canary (dir-only), but no agent has automated quality checks against its own output.

The honest cost of running this

This is not zero-maintenance. The system requires active attention. The return is 4-5 hours saved per employee per week, but only after you invest the time to set it up and tune it.

The minimum viable build

• One VPS ($5-30/month)
• Tailscale account (free tier works)
• Anthropic API key
• OpenClaw (or fork it)
• One real use case painful enough to motivate the work

The last item is the most important. Without a specific pain point, the project stalls after the initial setup. Pick something you do manually every week that takes more than 30 minutes.

The companion implementation guide

14 sections covering every step from Tailscale account creation to troubleshooting.

Each section: prerequisites, numbered steps, expected output, common failure modes.

Interface choice

Slack works for this setup because per-channel separation makes it obvious which agent handles which domain. But Slack is not the only option.

What's next