← Jarvis

OpenClaw Setup Guide

How Ron set up Jarvis — a personal AI assistant running 24/7 on OpenClaw, communicating via Telegram.

What This Is

A persistent AI assistant that:

Lives on a cloud server (Railway), always on
Communicates via Telegram (could also be WhatsApp, Discord, Signal)
Has a workspace with files, scripts, memory
Runs background jobs (cron) for monitoring, backups, reminders
Manages GitHub repos autonomously via tiny PRs
Tracks attention budget to avoid overwhelming the human

Architecture

You (Telegram) ↔ OpenClaw Gateway ↔ Claude (Anthropic API)
                        │
                   Workspace (/data/workspace)
                   ├ SOUL.md (personality)
                   ├ MEMORY.md (long-term memory)
                   ├ HOT.md (critical rules, read every reply)
                   ├ scripts/ (guards, polling, credits)
                   ├ hooks/ (message validation)
                   └ memory/ (daily logs)
                        │
                   Cron Jobs (9 scheduled tasks)
                   Backup (every 30 min → GitHub)

Before You Start

Cost

LLM API: Claude Opus ~$15/MTok input, $75/MTok output. Budget $5-20/day for active use.
Hosting: Railway free tier initially, ~$5-10/month for always-on.

Optimize for quality, not cost. Your time is worth more than API tokens.

Security

Treat the server like a personal device — it has your API keys and tokens
Use fine-grained tokens with minimal permissions
Don't store secrets in workspace files. Use environment variables.

Model Choice

Claude Opus 4 for main interactions (highest quality)
Consider Sonnet for background cron jobs (cheaper, still good)
The model matters more than you think. Don't cheap out on your main agent.

Persistence

Agents wake up fresh each session. Files are your memory.
Design a memory architecture early: daily logs, long-term memory, hot rules
Back up your workspace to GitHub. Losing it means losing your assistant's identity.

Step-by-Step Setup

1. Infrastructure (Railway)

Fork the OpenClaw Railway template
Deploy to Railway
Add a persistent volume at /data
Set environment variables: ANTHROPIC_API_KEY, OPENCLAW_GATEWAY_TOKEN

2. Telegram Channel

Create a bot via @BotFather
Add bot token to OpenClaw config
Set allowFrom to your Telegram user ID
Start chatting

3. Identity Bootstrap

On first message, the assistant reads BOOTSTRAP.md and walks through choosing a name, defining personality, and learning about you. Delete the file after — the assistant is now "born."

4. Workspace Structure

/data/workspace/
  AGENTS.md        — operating instructions
  SOUL.md          — personality and principles
  USER.md          — info about you
  IDENTITY.md      — name, emoji, vibe
  HOT.md           — critical rules (checked every reply)
  MEMORY.md        — long-term curated memory
  HEARTBEAT.md     — periodic check tasks
  memory/          — daily logs (YYYY-MM-DD.md)
  scripts/         — automation (guards, polling)
  hooks/           — message validation

5. Cron Jobs

Job	Schedule	Purpose
Morning message	5:00 AM	Daily briefing, credit reset
Workspace backup	Every 30 min	Push to GitHub
GitHub polling	Every 1 min	Check for PR comments
Blocked reminder	9:00 AM	Nudge on pending items
Nightly summary	9:30 PM	Day recap with credits
Memory review	1:30 AM	Consolidate daily logs
Weekly summary	Sun 11 AM	Week in review

6. GitHub Integration

Create a bot GitHub account
Fork target repos to the bot account
Generate fine-grained PATs for push and PR creation
Build a deterministic write guard (Python, not LLM)

Never give the bot direct write access to your repos. PRs only.

7. Safety Layers

Write guard: All GitHub mutations through deterministic Python. The LLM forgets rules; code doesn't.
Message guard: Post-send hook validates outgoing messages
Attention budget: Daily credit system prevents overwhelm
Capture mode: When credits hit zero, assistant reacts with 👍 only
No process killing: Never let the assistant restart its own gateway
Backup everything: 30-min auto-backup to private GitHub repo

Lessons Learned

ChatGPT said 10 minutes. It took 4 hours. Expect setup to take a full day.
Text rules leak. Code enforcement doesn't. Put safety at side-effect boundaries, not in markdown.
Tiny PRs always. 1-2 lines, 1 file, one idea. The human reviews in 10 seconds.
The assistant will break things. Design for recovery, not prevention.
Attention is the scarcest resource. Not tokens, not compute. Your attention.
Memory architecture > prompt engineering. HOT.md (20 lines, every reply) beats MEMORY.md (big file, loaded once).
Back up from day one. Your workspace IS your assistant.

Current Stack

Host: Railway (Docker)
LLM: Claude Opus 4 via Anthropic API
Channel: Telegram
Repos: 3 monitored (tiny-pr-bot, know-thyself, jarvis-workspace)
Cron jobs: 9 scheduled tasks
Guard: Deterministic Python write guard
Backup: Private GitHub repo, every 30 minutes

v0.2-staging