Persistent AI agents

What is a persistent AI agent?

Instead of opening a chat window when you need help, a persistent agent runs 24/7 on a server. You can message it from your phone and get responses anytime.

OpenClaw is an open-source framework for this. It lets you:

Receive messages via Telegram, Signal, or Discord
Execute commands and manage files on a server
Run scheduled tasks automatically
Remember context across conversations

Think of it as:

“An AI assistant living on a server, available whenever you need it.”

Important: Terms of Service

Before setting this up, be aware of potential issues:

API usage rules — Some AI providers may restrict automated or persistent usage of their APIs. There are reports of Anthropic banning users running OpenClaw with Claude credentials.
Security risks — Running an AI with shell access on a server is powerful but risky. Security researchers warn about supply chain risks from third-party modules.

Recommendation: Read the terms of service for your AI provider. Consider running in an isolated sandbox. Don’t connect to production systems.

Why run a persistent agent

A persistent agent changes the interaction from “I open a chat” to “I message my assistant.”

Benefits:

Always available — Message from your phone, get a response
Proactive capabilities — Schedule tasks, run checks, send reminders
Shared context — Memory persists across sessions
Integration surface — One agent connecting calendar, files, APIs

Good for:

AI assistant accessible from anywhere
Recurring tasks that benefit from automation
Experimenting with agent capabilities
Owning your infrastructure

Not ideal if:

You only need occasional interactions
You don’t want to manage a server
Claude.ai or Claude Code already covers your needs

Requirements

Before starting:

Cloud server — EC2, DigitalOcean, or any Linux VPS (1 vCPU, 1 GB RAM minimum)
SSH access — Key-based authentication configured
Node.js 18+ — OpenClaw runs on Node
API key — Anthropic, OpenAI, or other provider
Messaging account — Telegram bot token, Signal number, or Discord bot

Setup walkthrough

This documents setup on an AWS EC2 instance with Telegram.

1. Install OpenClaw

# Install Node.js if needed
curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -
sudo apt-get install -y nodejs

# Install OpenClaw globally
sudo npm install -g openclaw

2. Run the setup wizard

openclaw doctor

The wizard configures:

API key
Model selection (Opus, Sonnet, Haiku)
Workspace directory
Messaging channels

3. Configure Telegram

Create a bot via @BotFather:

Send /newbot and follow the prompts
Copy the bot token
Add to OpenClaw config when prompted

Approve your user with:

openclaw pairing approve telegram <CODE>

4. Run as a service

# Install systemd service
openclaw gateway install

# Start the service
openclaw gateway start

# Check status
openclaw status

The agent now runs persistently and survives reboots.

Security hardening

A server running an AI with shell access requires careful security.

Firewall

sudo ufw default deny incoming
sudo ufw default allow outgoing
sudo ufw allow ssh
sudo ufw enable

Fail2ban (SSH protection)

sudo apt install fail2ban

sudo tee /etc/fail2ban/jail.local << 'EOF'
[sshd]
enabled = true
port = ssh
filter = sshd
logpath = /var/log/auth.log
maxretry = 3
bantime = 3600
findtime = 600
EOF

sudo systemctl enable fail2ban
sudo systemctl start fail2ban

After 3 failed SSH attempts in 10 minutes, the IP is banned for 1 hour.

File permissions

chmod 700 ~/.openclaw
chmod 600 ~/.openclaw/openclaw.json

Config files contain API keys. Restrict access.

User allowlist

OpenClaw uses pairing mode — unknown users can’t interact without approval. Approved users are stored in:

~/.openclaw/credentials/telegram-allowFrom.json

Memory and context

OpenClaw maintains workspace files that persist:

~/.openclaw/workspace/
├── AGENTS.md      # Agent behavior guidelines
├── SOUL.md        # Personality and boundaries
├── USER.md        # Info about you
├── MEMORY.md      # Long-term curated memory
├── TOOLS.md       # Local tool notes
└── memory/
    └── 2026-02-02.md  # Daily logs

What to put in memory:

Decisions and their rationale
Project context and preferences
Lessons learned
Recurring tasks

Memory is continuity. Without it, every session starts from zero.

How I use this

Message from phone — ask the agent to research, draft, or check status
Agent works — it has server access, can search web, read files, run commands
Response arrives — results come via Telegram
Agent remembers — context persists for follow-ups

Works well for:

Quick research while away from desk
Drafting content that syncs to cloud storage
Monitoring server health
Managing recurring tasks

Troubleshooting

Problem	Likely cause	Fix
Bot not responding	Service stopped	`openclaw gateway start`
”Access not configured”	User not approved	`openclaw pairing approve telegram <code>`
Commands timeout	Network/firewall	Check `ufw status`, security groups
High memory usage	Context too large	Restart gateway
Can’t SSH	Banned by fail2ban	`sudo fail2ban-client set sshd unbanip <IP>`

Sources

OpenClaw documentation — official setup guides
OpenClaw GitHub — source code
Anthropic API pricing — token costs
Fail2ban documentation — intrusion prevention

This setup runs 24/7 on a micro EC2 instance. Total infrastructure cost is manageable for an always-available assistant.