Skip to main content
Mythos

OpenClaw is an open-source personal AI assistant platform that runs as a local agent on a user's own hardware, executing tasks autonomously across messaging channels and the local system. Originally built as Clawdbot by Peter Steinberger and renamed OpenClaw in early 2026, it connects agents to 📝WhatsApp, 📝Telegram, 📝Discord, and Slack, and routes work across models from 📝Anthropic, 📝OpenAI, and local installs. Its defining move is to put an agent's soul, identity, tools, and memory in plain markdown files you can read, version, and edit by hand.

Why local, file-defined agents matter

OpenClaw arrived in January 2026 into a market dominated by cloud-hosted agent products, and took the opposite bet: the agent runs on hardware you own, reachable through the chat apps you already use. That bet changes who controls the system. A cloud agent is a tenant on someone else's platform; an OpenClaw agent is software on your own machine, its behavior defined in files you can diff. As personal AI shifts from novelty to infrastructure, the question stops being which model is smartest and becomes who owns the agent that runs your day. OpenClaw is one of the first platforms to answer plainly: you do.

The markdown-file agent model

OpenClaw's most distinctive idea is that an agent is a set of markdown files, not a system prompt. Each agent is shaped by SOUL.md (temperament and voice), IDENTITY.md (role and boundaries), TOOLS.md and SKILLS.md (capabilities), MEMORY.md and USER.md (long-term context), HEARTBEAT.md (scheduled, autonomous activity), and BOOTSTRAP.md (initialization). Because identity lives in files, agents can be version-controlled, templated, and evolved like any codebase — you change behavior by editing a file and create a new agent by copying one. The file-defined model, its seven-surface memory layout, and how its capability surfaces differ from MCP are detailed in the cluster memos below.

Orchestration, channels, and memory

Beyond identity, OpenClaw is an orchestration layer: it routes tasks to the right agent, spawns child sessions for 📝multi-agent orchestration, manages connections across channels, and directs work to the appropriate model with fallback chains. A plugin architecture extends it, cron jobs schedule autonomous runs, and a browser pool handles web automation. Through 📝MCP integration, agents reach external systems — including a 📝MythOS library used as shared, persistent memory across every agent: OpenClaw supplies the Mind, MythOS the Memory.

Where OpenClaw fits among agent frameworks

OpenClaw occupies a different niche than the developer-facing agent frameworks. Where CrewAI, LangChain, and AutoGPT are libraries you assemble into an application, OpenClaw is a running, multi-channel personal assistant you configure with files. The full comparison against those frameworks — and where file-defined agents sit inside the broader practice of human-AI augmentation — is covered in the cluster below.

Related

For about ten years I avoided learning to code on purpose — I knew myself well enough to know I'd vanish down the rabbit hole. OpenClaw is what finally got me to relent. I ran it hard, and for a while it was the whole nervous system of my augmented self. Then the friction set in: I was forever accommodating its hiccups as it drove MythOS through browser control. One day, fed up — and by then deep in 📝Claude Code — I pointed an agent swarm at the problem and rebuilt MythOS from scratch. That rebuild is the MythOS I run today, and it's how I became its engineer. OpenClaw was the door; I walked through it and kept going.

Contexts

Created with 💜 by One Inc | Copyright 2026