The open-source always-on agent runtime — an OS for a personal AI employee that you deploy to a VPS, design with md files, and talk to over Telegram.

Pick this when

You want a proactive 24/7 assistant (schedules, heartbeats, channel inboxes) that YOU design and host — not another chat app you have to open.

What makes it different

Proactive, not reactive: triggers (timers, calendar events, heartbeats, messages) make the agent come to you

The md-file design surface — user/identity/soul/memory + tools/heartbeat/bootstrap — turns setup into agent DESIGN

Runs anywhere (local, Hetzner, Hostinger 1-click, Railway template) with model routing across providers

What it does

Runs an always-on gateway that routes Telegram/WhatsApp messages, cron-style heartbeats, and skills

Reads, edits, and creates files in its workspace — including its own instruction documents

Routes models per task (reasoning vs summary vs chit-chat) when you configure it to

What it does not do

Common wrong expectations

Isn't an agent out of the box — 'when you install OpenClaw, there's no agent'; you design one on the runtime

Doesn't guard itself — security (fresh accounts, allowlists, read-only-first, kill switch) is entirely on the builder

Doesn't conserve tokens by default — heartbeats, memory decay, and premium default models all burn

Where it sits in a stack

The archive's central case study in agent INFRASTRUCTURE thinking — the mobile-phone analogy (an OS is useless until you add SIM and apps) that also explains Hermes, harness folders, and CLAUDE.md.

Shelf

Automation platforms · Chat assistants

Our status · watch

Two full sprint sessions of doctrine; the md-file architecture matters to Paul more than the product itself.

For your projects

The user/identity/soul/memory split is the same behavior-vs-state architecture as Paul's hard-rules + open-items tracker.

Freshness

fast-moving

Related tools

complementsClaude CodecomplementsHostinger / SiteGroundcompetes withn8n

claude-code: Suken drives OpenClaw server changes THROUGH Claude Code; Claude models do its heavy reasoning

hostinger: The sprint's beginner path: 1-click managed OpenClaw on a Hostinger VPS

n8n: 'n8n is automations; OpenClaw is autonomous' — his line for where each wins

Every moment the course touched it (6)

SessionCoverageMomentWhat happened
AI Sprint: OpenClaw — Session 1 (Infrastructure, Not Agent: the Seven Documents, 1-Click VPS, and Model Routing)demonstratedThe subject: runtime installed via Hostinger 1-click, designed live with documents and a pasted skill
AI Sprint: OpenClaw — Session 2 (Token Anatomy, Claude Code Inside the Server, Skills, and Multi-Agent Workspaces)demonstratedSelf-managed on VPS: token anatomy, skills panel, multi-agent workspaces, the Duma lead agent
Session 14: Claude Code — Advanced Workflowsmentioned1:19:23Formerly 'Clawd' until Anthropic's trademark action; its ClawdHub was the injection incident
Session 15: Loop, Harness & Graph Engineering (posted as 'Securing Vibe-Coded Apps')mentioned0:40:04Daily zero-cost driver with Hermes on GLM 5.2 (as heard): sheets, bookings, Meet links; also credited with waking the market up to agent memory
AI Masterclass: Exploring Antigravity 2 — Four Products, Four Slash Commands, and Two Live BuildsmentionedSame autonomous-agent category contrast; swarm comparison later
AI Sprint: Windsurf & Devin — Day 1 (Vibe to Agentic: GitHub as Pizza, the Four Spec Concerns, and the IDE Lineage)mentionedThe rate-limit case study: 'Anthropic rate-limited OpenClaw'

← All tools by category