Agent
An AI-powered assistant that runs on your machine, reads instructions from your workspace files, and executes tasks using tools. In OpenClaw, each agent has its own session, memory, and personality.
What is OpenClaw?AGENTS.md
The main instruction file in your workspace. It tells the agent how to behave, which files to read on startup, and what safety boundaries to respect. Loaded at the start of every session.
Setup guideAllowlist
A list of approved contacts or phone numbers that are permitted to message your agent. Used in WhatsApp configuration to prevent unauthorized access. If you skip this, anyone who texts the agent number gets full access.
Security guideAnsible Installer
An automated installation method for deploying OpenClaw on remote servers. Runs a playbook that handles dependencies, configuration, and systemd service setup in one command.
Installation guideAPI Key
A secret string that authenticates your agent with an AI model provider like Anthropic or OpenAI. Stored in environment variables or your OpenClaw config file. Never commit these to version control.
Security guideBootstrap
The first-run ritual that happens when a new agent starts for the first time. The agent reads BOOTSTRAP.md, asks you introductory questions, creates IDENTITY.md, and then deletes the bootstrap file so it only runs once.
Setup guideBridge Protocol
The communication layer between the Gateway and external channels like Telegram or WhatsApp. It translates incoming messages into a format the agent can process, and sends agent responses back to the correct channel.
Browser Tool
A built-in OpenClaw tool that gives the agent the ability to open web pages, click elements, fill forms, and extract content from websites. Runs in a headless browser on your machine.
Channel
A messaging platform connected to your agent. Telegram, WhatsApp, and Discord are the three supported channels. Each channel gets its own session context so conversations stay separate.
Channel setupClawHub
The community marketplace for sharing and discovering OpenClaw skills. Browse pre-built automations, install them into your workspace, and publish your own for others to use.
CLI
Command Line Interface. The primary way to interact with OpenClaw outside of chat. Used to start the Gateway, manage channels, configure settings, and run diagnostics. The main command is "openclaw".
Installation guideCompaction
When a conversation gets too long for the model's context window, OpenClaw compresses older messages while preserving key facts. Before compacting, the agent writes durable notes to disk so nothing critical is lost.
Context Window
The maximum amount of text an AI model can process in a single conversation. When the conversation exceeds this limit, OpenClaw triggers compaction to free up space while saving important context to memory files.
Cron Job
A scheduled task that runs on a fixed interval using your operating system's scheduler. OpenClaw can be configured to run specific automations on a schedule, separate from the heartbeat system.
DM (Direct Message)
A private one-on-one conversation with your agent, as opposed to a group chat. In OpenClaw, DMs use the "main session" which has access to MEMORY.md and full long-term context.
Docker
A containerization platform that packages OpenClaw and its dependencies into an isolated environment. Recommended for production deployments because it keeps your host system clean and makes updates simple.
Docker setupDocker Sandbox
A security feature that runs agent tool execution inside a Docker container. Prevents the agent from accessing your host filesystem or running commands outside the container. Configurable per session type.
Docker setupElevated Mode
A special permission level that grants the agent access to restricted tools or actions. Must be explicitly enabled in configuration. Used when you trust the agent to perform sensitive operations like file deletion or system commands.
Exec Tool
A built-in OpenClaw tool that lets the agent run shell commands on your machine. Subject to tool policy restrictions. Can be sandboxed to limit what commands are allowed.
Failover (Model Failover)
Automatic switching to a backup AI model when the primary provider is down or rate-limited. Configure a primary model and one or more fallbacks in your OpenClaw config. The switch happens silently with no user action required.
Formal Verification
A mathematical approach to proving that software behaves correctly. OpenClaw uses TLA+ specifications to verify critical system properties like message ordering and session isolation.
Gateway
The core server process in OpenClaw. It manages agents, routes messages between channels, handles authentication, and serves the web UI. Started with "openclaw gateway start" and runs as a background service.
Setup guideGateway Token
An authentication token that protects the Gateway API when it is exposed on a network. Required when running on a VPS or any non-localhost setup. Set via the OPENCLAW_GATEWAY_TOKEN environment variable.
VPS setupHeartbeat
A periodic check-in where the Gateway prompts the agent to run a checklist of tasks. Default interval is 30 minutes. The agent reads HEARTBEAT.md, performs the listed actions, and reports back. Each heartbeat costs one API call.
Heartbeat setupHEARTBEAT.md
A workspace file containing a short checklist of tasks the agent should perform during each heartbeat. Keep it concise to control API costs. Example tasks: check email, review calendar, monitor feeds.
Setup guideIDENTITY.md
A workspace file created during the bootstrap ritual. Contains the agent's chosen name, emoji identity, and personality vibe. Auto-generated during first run, but you can edit it afterward.
Setup guideLLM (Large Language Model)
The AI model that powers your agent's reasoning and responses. OpenClaw supports models from Anthropic (Claude), OpenAI (GPT), and local providers like Ollama. You choose which model to use in your config.
Local Model
An AI model running directly on your machine instead of through a cloud API. Supported via Ollama integration. Slower than cloud models but keeps all data on your hardware with no API costs.
MCP (Model Context Protocol)
An open standard for connecting AI models to external tools and data sources. OpenClaw supports MCP servers, letting you plug in tools from the broader MCP ecosystem without custom integration code.
MEMORY.md
An optional workspace file for curated long-term memory. Contains important facts, preferences, and decisions that should persist across weeks and months. Only loaded in main sessions (direct messages) for security.
Memory systemModel Provider
A service that hosts and serves AI models. Anthropic, OpenAI, and Google are cloud providers. Ollama is a local provider. You can configure multiple providers and set failover order in OpenClaw.
Multi-Agent
Running more than one agent instance, each with its own workspace, personality, and channel assignments. Useful for separating work and personal assistants, or giving different teams their own agent.
Node
In the context of OpenClaw deployment, a single server or machine running the Gateway process. In multi-node setups, agents can be distributed across machines for load balancing and redundancy.
OAuth
An authentication protocol used by many third-party services (Google, GitHub, Slack). OpenClaw stores OAuth tokens in ~/.openclaw/credentials/ so your agent can access connected services without re-authenticating.
Ollama
An open-source tool for running AI models locally on your own hardware. OpenClaw integrates with Ollama as a model provider, giving you a free, private alternative to cloud APIs.
Onboarding Wizard
An interactive CLI walkthrough that helps you configure OpenClaw for the first time. Guides you through model selection, channel setup, and security settings. Run with "openclaw configure".
Installation guideOpenClaw
An open-source AI agent platform that runs on your own machine. It connects to messaging channels, executes tasks using built-in tools, and maintains persistent memory across sessions. Self-hosted, privacy-first, and fully customizable.
Full overviewPairing
The process of connecting two agents so they can communicate and collaborate on tasks. Paired agents can delegate work to each other and share context within defined boundaries.
Plugin
An extension that adds new capabilities to OpenClaw. Plugins can provide tools, channels, or integrations with external services. Installed via the CLI or from ClawHub.
Prompt Injection
A security attack where someone tricks the agent into executing unintended instructions by embedding commands in seemingly normal text. OpenClaw mitigates this with tool policies, sandboxing, and session isolation.
Safety overviewProvider
See Model Provider. The AI service that supplies the language model your agent uses for reasoning. Configured in the models.providers section of openclaw.json.
Sandbox
An isolated environment that restricts what the agent can access on your machine. Docker-based sandboxing prevents file system access and limits shell commands. Configurable per session type (main, group, sub-agent).
Security guideSeccomp Profile
A Linux kernel feature that restricts which system calls a sandboxed container can make. OpenClaw applies a seccomp profile to Docker sandbox containers for an additional layer of security beyond filesystem isolation.
Session
A single conversation thread between you and the agent. Each session has its own context, loaded workspace files, and message history. New sessions start fresh. Use /new in chat to begin one.
Skill
A reusable automation template that teaches the agent how to perform a specific task. Skills live in your workspace's skills/ directory as Markdown files with instructions, examples, and tool configurations.
SOUL.md
A workspace file that defines the agent's personality, tone, and communication boundaries. Think of it as a character sheet. Read every session to maintain consistent behavior. Keep it under 500 words to save tokens.
Setup guideSub-Agent
A secondary agent spawned by the main agent to handle a specific subtask. Sub-agents run in isolated sessions with limited permissions. The main agent delegates work and receives results back.
Systemd
A Linux service manager that keeps OpenClaw running in the background and restarts it if it crashes. The recommended way to run OpenClaw on a VPS. Configured automatically by the Ansible installer.
VPS setupTailscale
A VPN service that creates a private network between your devices. Recommended for securely accessing your OpenClaw Gateway from outside your local network without exposing ports to the public internet.
VPS setupTLA+ (Formal Verification)
A specification language used by OpenClaw's developers to mathematically verify system correctness. Ensures properties like "messages are never delivered out of order" and "sessions are always isolated."
Tool
A capability the agent can use to interact with the outside world. Built-in tools include shell execution, file editing, web browsing, and HTTP requests. Custom tools can be added via MCP or plugins.
Tool Policy
Rules that control which tools the agent is allowed to use and under what conditions. Configured in openclaw.json under tools.sandbox.tools. You can allow, deny, or require confirmation for specific tools.
Security guideUSER.md
A workspace file that tells the agent who you are. Your name, timezone, preferences, and how you like to be addressed. Prevents the agent from asking basic personal questions every session.
Setup guideVPS (Virtual Private Server)
A remote server you rent to run OpenClaw 24/7. Providers like DigitalOcean, Hetzner, and AWS offer affordable VPS options. Running on a VPS means your agent is always online, even when your laptop is off.
VPS setup guideWebhook
An HTTP endpoint that receives event notifications from external services. OpenClaw can expose webhooks so third-party tools (Stripe, GitHub, etc.) can trigger agent actions when something happens.
WebChat
The built-in browser-based chat interface for talking to your agent. Available at the Gateway's local address (default: localhost:18789). Works immediately with no channel setup required.
Workspace
The directory (~/.openclaw/workspace/) that contains all of your agent's instruction files, memory, and skills. This is the agent's brain. Back it up with git. Keep it separate from the config directory.
Workspace walkthroughDefinitions by The Operator Vault team, based on months of hands-on OpenClaw experience. Found an error or missing term? Let us know in the community
Written by
Kevin Jeppesen
Founder, The Operator Vault
Kevin is an early OpenClaw adopter who has saved an estimated 400 to 500 hours through AI automation. He stress-tests new workflows daily, sharing what actually works through step-by-step guides and a security-conscious approach to operating AI with real tools.
Glossary FAQ
Questions about terminology.
Ready to get started?
Learn by doing.
The $19 workshop covers everything in this glossary, hands-on. Install, configure, connect, and run your first workflow in one sitting.