41 terms, plain English
OpenClaw Glossary
Every OpenClaw term in one place. The .md files that govern how an agent thinks and acts, the channels it can talk through, the internals you'll only see once you open the source, and the adjacent tools that show up in almost every real deployment. Each entry is one or two sentences, no jargon for the sake of jargon.
Authored by Adhiraj Hangal, founder of OpenClaw Consult and a merged contributor to openclaw/openclaw core.
Files (the .md system)
OpenClaw treats agent behaviour as a stack of plain-text Markdown files. Each one owns a different layer of how the agent thinks, remembers, and acts.
- AGENTS.md
- The operating manual for the agent. Defines hard rules, safety defaults, sub-agent delegation, tool policies, and the security mechanisms that make autonomy safe.
- HEARTBEAT.md
- The proactive automation file. The agent reads this on a recurring tick and decides whether to act on its own, send a morning briefing, run a scheduled task, check a system, all without being prompted.
- SOUL.md
- Where the agent's personality lives. Tone, opinions, boundaries, voice. Without SOUL.md, every agent sounds like generic AI. With it, the agent has a recognisable point of view.
- MEMORY.md
- The long-term memory file. Holds durable facts the agent should always know about the user, the business, and prior conversations. Pairs with daily notes and a dreaming system that promotes what matters.
- SKILLS.md
- Index of installed Skills, the modular packages that give the agent new capabilities. Each Skill is a folder with its own instructions, scripts, and resources, loaded on demand.
- TOOLS.md
- Declares which low-level tools the agent is allowed to call, things like shell exec, file read, file write, web fetch, browser. Pairs with the policies in AGENTS.md for safety.
- KNOWLEDGE.md
- Reference material the agent can quote verbatim, product docs, pricing, policies, the kind of facts you want answered word-for-word, not paraphrased.
Channels
A channel is the surface a user actually talks to the agent through. One agent can be wired into many channels at once, and each channel handles formatting and access control on its own terms.
- Telegram channel
- Connects the agent to a Telegram bot. The fastest channel to set up, supports markdown, inline buttons, and group chats. Strong default for a personal-use agent.
- WhatsApp channel
- Connects the agent to a WhatsApp Business number. Best when the user is non-technical or already lives in WhatsApp. Requires a Twilio or Meta API account, longer setup than Telegram.
- Discord channel
- Connects the agent as a Discord bot. Good for community-facing agents, supports threads, slash commands, and per-channel scoping.
- iMessage channel
- Connects the agent to iMessage on a Mac. Native blue bubbles, no third-party app required. Mac-only on the agent side, the user is anyone with an iPhone.
- Slack channel
- Connects the agent to a Slack workspace. Best for team-facing agents, integrates with workspace permissions and supports per-channel routing.
- WebChat channel
- The built-in browser chat surface. No third-party signup needed, the agent is reachable through a web URL. Useful for embedding inside a customer-facing site.
- Email channel
- Connects the agent to a mailbox so it can read inbound messages and reply. Best for asynchronous workflows where Telegram or WhatsApp would feel too noisy.
Concepts
The high-level vocabulary that shows up across every OpenClaw doc, channel guide, and lab post.
- OpenClaw
- The open-source agentic AI framework created by Peter Steinberger. Lets a single agent run autonomously, use tools, send messages, write and execute code, and complete multi-step tasks without human intervention.
- OpenClaw runtime
- The Node.js process that executes an OpenClaw agent. Loads the .md files, opens the channel connections, and runs the agent loop. Lives in the openclaw/openclaw repo.
- OpenClaw agent
- A single configured instance of the runtime. One agent has one identity, one set of .md files, and one set of channels. Multi-agent setups run several agents behind a shared gateway.
- AI provider
- The model backend the agent talks to. Anthropic, OpenAI, Google, and local Ollama are all supported. Most production builds use a two-tier setup, a fast cheap model for routine work and a smarter expensive model for hard prompts.
- Tool
- A low-level capability the agent can invoke, things like exec, read, write, edit, web fetch, browser navigation. Tools are the smallest unit of action and are governed by AGENTS.md.
- Skill
- A modular package that bundles instructions, scripts, and resources to give the agent a new capability. Skills are how you teach the agent to do something specific, like draft a quote or post to a CRM.
- Channel
- The user-facing surface for an agent, Telegram, WhatsApp, Discord, iMessage, Slack, WebChat, Email. One agent can be connected to many channels at once.
- Heartbeat
- The recurring tick that lets an agent act without being prompted. The agent checks HEARTBEAT.md on every tick and decides whether to do something. Heartbeat math is one of the biggest cost levers in production.
- Memory
- Anything the agent can remember across sessions. OpenClaw splits memory into MEMORY.md for long-term facts, daily notes for running context, and a dreaming system that promotes what matters.
- Soul
- The agent's voice and behaviour. Shaped by SOUL.md, the Molty prompt, and the operating rules in AGENTS.md.
- Knowledge
- Reference material the agent can quote verbatim, sourced from KNOWLEDGE.md and any attached docs. Distinct from memory, which is about prior interactions.
Internals
The names you see when you open the openclaw/openclaw source, debug a stalled agent, or read the changelog. Useful for engineers wiring the runtime into a production deployment.
- pi-agent-core
- The core agent package inside openclaw/openclaw. Owns the agent loop, the tool dispatcher, and the contract every channel and Skill plugs into.
- pi-embedded-runner
- The wrapper that runs an OpenClaw agent inside another process, embedded mode. The outer agent run loop and the idle-timeout circuit breaker added in PR #76345 live here.
- Idle timeout
- The signal the runtime emits when an LLM call returns no output before the deadline. A single idle timeout is normal, repeated idle timeouts in a row mean the connection is wedged and the agent should stop retrying.
- Circuit breaker
- A safety mechanism that stops trying after N consecutive failures. In OpenClaw, the cost-runaway breaker added in PR #76345 caps consecutive idle timeouts at 5, so a wedged provider can no longer fan paid model calls forever.
- ClawHub
- The OpenClaw Skills marketplace. Where third-party developers publish Skills and where you install them from. Everything you install carries security risk, treat ClawHub Skills the same way you'd treat an npm package.
- MoltBook
- The reference handbook that ships alongside the runtime, where the canonical bootstrap order, config keys, and security audit checklist live. The day-16 lesson of the bootcamp walks through it end-to-end.
- ZeroClaw
- The minimal OpenClaw distribution, agent loop and channels only, without Skills, ClawHub, or the dreaming system. Used for low-resource hardware and stripped-down embedded deployments.
Operations
What you actually buy when you hire someone to do OpenClaw work for you.
- OpenClaw consulting
- Scoping, architecture, and design work for an OpenClaw deployment. Usually billed before any engineering starts so the scope and timeline are agreed in writing first.
- OpenClaw implementation
- Building the actual agent. Configuring the .md files, wiring channels, writing custom Skills, integrating with the rest of the business stack. Most production builds at OpenClaw Consult take 2-4 weeks.
- OpenClaw maintenance
- Ongoing work after launch, monitoring, fixing what breaks, applying upstream updates, evolving the agent as the business changes. Usually a monthly retainer.
- OpenClaw bootcamp
- The free 16-day OpenClaw course at openclawconsult.com/bootcamp. 3+ hours of expert video instruction taught by Adhiraj Hangal, covering architecture, install, models, channels, every .md file, Tools and Skills, Docker, and production deployment.
Adjacent
Tools and platforms that show up in almost every real OpenClaw deployment.
- Ollama integration
- Running OpenClaw against a local Llama 3 or similar open-weight model via Ollama. Zero API cost, full data privacy, slower than cloud providers and capped by the host's hardware.
- Anthropic Claude integration
- OpenClaw's most-used cloud provider. Claude Sonnet for the smarter tier, Haiku for the cheap routine tier. The cost-runaway bug fixed by PR #76345 was first observed against Claude Sonnet 4.6.
- OpenAI integration
- GPT-class models as an OpenClaw provider. Useful for tool-use heavy workflows and as a fallback when Anthropic is rate-limited.
- Docker deployment
- Running OpenClaw inside a Docker container so it's isolated from the host and reproducible across machines. Day 12 of the bootcamp covers Docker Compose and sandbox modes end-to-end.
- Vercel deployment
- Hosting the web surface around an OpenClaw deployment, dashboards, WebChat embeds, marketing pages, on Vercel. The agent itself usually runs on a VPS or container, not on Vercel's serverless runtime.
Spotted a missing term, or a definition that needs a sharper edge? Email experts@openclawconsult.com and it'll get added.