1 / 10
ERA · March 2026

Getting Started
with OpenClaw

or: How I Spent My Weekends Arguing with a Config File

Chuck Blake

Background

What is OpenClaw?

  • Local AI agent framework — runs on YOUR machine
  • Connects to Telegram (or Signal, Discord, etc.)
  • Always-on personal assistant with memory, tools, and actions
  • Think: chief of staff, not a chatbot
  • Why local: your data, your vault, your rules
Setup

Getting Started

30-minute path: Install → Telegram bot → Connect → Write SOUL.md → First conversation

⚠️ Hidden gotcha: default idle timeout = 2 hours → change to 4 days immediately
thread: { mode: "idle", idleMinutes: 5760 }
Architecture

Memory Architecture

Files
IDENTITY.md · errors.md · MEMORY.md · Daily log
Vector Store
Mem0 + Qdrant (Docker)
Boot retrieval
Recency pull — 20 most recent memories
Per-question
Semantic search by topic
Embeddings
nomic-embed-text (100% local)
LLM
qwen3.5:9b via Ollama
Security

Secrets Management

  • Gomez has his own 1Password vault — scoped, can't see personal vault
  • Service account token (read-only)
  • agent-vault = convenience cache layer
  • Least privilege, scoped, no shared credentials
Components

Skills + Scripted Components

Skills = instruction sets  ·  Scripts = reliable primitives

Scripts: morning_briefing.py  ·  health_log.py  ·  omnifocus_reset_planned.sh  ·  op-wrapper.sh

Skills: good-morning  ·  deploy  ·  pdad-do-next  ·  omnifocus-process-inbox

"Agent orchestrates — scripts execute."
Access Model

How Much Access?

🟢
Read-only bus (start here)
Production events → Slack → agent reads only
🟡
Read-only production
Log access only, can analyze
🔴
Full scoped access
GitHub App per repo, Heroku, 1Password vault
"Start small. Expand after trust."
Field Notes

Lessons Learned

  1. 1Start small — many cycles of tweaking
  2. 2Scripts for consistent output
  3. 3Programmatic enforcement > prompt guidance
  4. 4Self-diagnosis to self-correction cycle
"If your system can't refine itself based on what went wrong last week, you're just running the same broken process faster."

Real failures: JSON crash  ·  ghost ✅  ·  sub-agent recreating deleted crons  ·  R14 flood

Roadmap

Where It's Going

  • Whistler: autonomous coding sub-agent
  • PDAD loop: mostly autonomous product development
  • Daily ops on autopilot
  • Vision: agent handles operational layer; I focus on judgment
Go Build

Resources

🦀
OpenClaw
openclaw.ai
 🔧
ClawHub — Skills & Plugins
clawhub.com
 📚
Anthropic Learning
anthropic.skilljar.com