OpenClaw

Open-source personal AI agent platform combining a local gateway, multi-agent orchestration, and channel-based messaging interfaces.

Last updated: 2026-04-24

Overview

OpenClaw is an open-source framework for running persistent, autonomous AI agents on owned hardware (Mac Mini, VPS, or hosted service). Unlike chat-first AI tools, OpenClaw agents operate on a continuous schedule — checking a heartbeat every 30 minutes and running cron jobs independently, even overnight. A user interacts with agents over messaging channels (Telegram, WhatsApp, Slack) rather than a browser interface.

Jensen Huang called it “probably the single most important release of software, probably ever.” It embodies Andreessen’s Unix agent architecture: LLM + bash shell + file system + markdown + cron job.

Alex Krantz (UC Berkeley PhD, Sky Lab) spent a month running OpenClaw in production and several weeks in the codebase, producing the most technically precise architectural analysis available. His framing: OpenClaw is “gross” code that would not pass Google review, but its design abstractions are elegant — and in the current era, design quality matters more than implementation quality.

Three-Layer Architecture (Krantz Analysis)

┌──────────────────────────────────────────────┐
│  Connectors (Telegram, WhatsApp, Slack, etc) │  ← human interface layer
├──────────────────────────────────────────────┤
│  Gateway Controller                          │  ← sessions, cron, heartbeat
│    Sessions = OS processes (isolated)        │
│    Cron & heartbeat = time management        │
├──────────────────────────────────────────────┤
│  Agent Runtime                               │  ← tools, skills, memory
│    Agents within sessions = threads          │
│    Intersession communication tools          │
└──────────────────────────────────────────────┘

Sessions as Processes, Agents as Threads

OpenClaw maps cleanly onto OS abstractions:

  • Sessions = OS processes — isolated context, separate permissions, can run in a sandbox. A Discord server becomes a natural session organizer: each channel = a separate session = isolated context.
  • Agents within a session = OS threads — they share session context but run independently.
  • Intersession communication: dedicated tools for agents in different sessions to pass messages, including agent-to-agent email (dedicated agent email address enables agent-to-agent skill sharing).

Heartbeat vs. Cron: Control Over Time

OpenClaw gives agents two distinct mechanisms for autonomous scheduling:

MechanismWhen to useHow it works
Heartbeat (30-min default)Unpredictable monitoring, reactive tasksFires heartbeat.md content + prior heartbeat history to the LLM; can trigger intersession messages
CronPredictable scheduled tasksStandard cron expression; fires at specified time

Together they give agents “control over time” — they can act without human prompting on both schedule-driven and event-driven triggers.

Configuration Files

Each agent workspace contains markdown files loaded at startup:

FilePurposeNotable detail
bootstrap.mdSelf-configuration on startup”You just woke up. Time to figure out who you are.”
soul.mdSelf-evolving persona”You’re not a chatbot. You’re becoming someone.” — instructed to update itself as the agent learns
user.mdProfile of the human operator
agents.mdSecurity and memory reminders
tools.mdTool usage tips and patterns

The soul.md self-modification design is the “strange loops” observation: the agent becomes the interface for reconfiguring itself through LLM calls. Krantz references Hofstadter’s Gödel, Escher, Bach as the conceptual parallel.

Skills: Three-Level Fidelity

OpenClaw implements a three-level lazy loading system for skills:

  1. Header (trigger) — always included in context by default; just enough to recognize when a skill applies
  2. Body (how-to) — fetched on demand when the agent identifies it may use the skill; full procedure
  3. Linked files (examples, scripts) — fetched only if the body is insufficient

This solves the context window problem: the system supports 150+ skills (capped at 150 or 30,000 characters), with smart filtering beyond that threshold. Skills are never front-loaded — only the headers are in context until needed.

Skills > MCP in practice: Krantz’s empirical observation after a month in production — MCP server adoption has faded because agents are now adept at CLI tools directly. Skills are easier to write, remarkably effective, and fit the thin-harness pattern better than MCP wrappers.

Lazy Memory Loading

OpenClaw uses a vector database for past conversation history, but does not front-load it into context. The agent uses memory-search and memory-get tools only when it decides it needs them. This avoids the “context stuffing” failure mode while keeping full history accessible.

Self-Extending Architecture

OpenClaw agents can:

  • Add their own plugins
  • Discover and install skills autonomously
  • Configure their own connectors through the UI

This is the strange-loops property made concrete: the agent reconfigures its own capabilities through the same LLM calls it uses for everything else.

Deployment Options

  1. exc.dev (recommended by Krantz) — $20/month, 50 persistent VMs, secure networking comparable to Tailscale; compute lives at LLM providers, no need for dedicated Mac Mini hardware
  2. Hosted — StartClaw, MyClaw, SimpleClaw, UniClaw, Every’s Plus One (slickest UX, improving fast)
  3. VPS — Railway, Hostinger, DigitalOcean, Google Cloud, Render (cheapest, most technical)
  4. Local hardware — Mac Mini M4 (~$600) or any spare laptop (most educational, requires isolated machine)

Critical safety note: Never install on a primary work or personal computer. Use an isolated machine — OpenClaw can access all files on its host.

Six Starter Workflows

Claire Vo’s recommended entry points:

  1. Weekend logistics — weekly family calendar coordination via group message
  2. Social media — trending topic research → meme generation → post approval → TikTok publish
  3. PLG sales prospecting — enrich new signups, segment by company size, auto-send sales emails
  4. Meeting prep — 30-min pre-brief via Telegram with attendees, agenda, and email history
  5. Support docs — weekly scan of resolved tickets → auto-create Linear issues for FAQ candidates
  6. Project management — maintain task list, break into daily actions, weekly celebration/flag

The Meta-Rule (Claw Operating Principle)

You are not allowed to do one-off work. If I ask you to do something that will happen again: do it manually on 3–10 items, show me the output, if approved codify it into a skill file, if it should run automatically put it on a cron. If I have to ask you twice, you failed.

This is a direct instantiation of thin-harness-fat-skills applied to a personal agent setup.

Connections

  • thin-harness-fat-skills — OpenClaw embodies this pattern: agents are thin harnesses; skills are fat and reusable; 3-level fidelity is an OpenClaw-specific implementation
  • agent-memory — workspace markdown files (bootstrap.md, soul.md) are OpenClaw’s memory layer; lazy vector DB avoids context stuffing
  • marc-andreessen — coined “Unix agent architecture” (LLM+shell+files+cron) that OpenClaw implements
  • claire-vo — wrote the definitive setup guide; runs 9 agents across work and personal life
  • alex-krantz — produced the definitive architectural analysis; loopiness model; “design > implementation” thesis
  • ai-agents — OpenClaw is a consumer-facing instantiation of agent-first software

Sources