A reference for all terms, concepts, and jargon used throughout this book and the OpenClaw ecosystem.
Agent An AI worker configured inside OpenClaw. An agent has a provider (Anthropic, OpenAI, etc.), a model, a system prompt, and a set of skills. When a user sends a message, the gateway routes it to an agent which processes it and returns a response.
Agent Provider The AI service that powers an agent. OpenClaw supports Anthropic (Claude), OpenAI (GPT), Google (Gemini), Ollama (local models), and others. Each provider has different models and pricing.
Allowlist
A list of user IDs or patterns that are permitted to interact with a workspace. Users not on the allowlist are blocked. The wildcard * allows anyone.
A2UI (Agent-to-UI) OpenClaw's protocol that allows agents to render UI components — cards, buttons, forms, tables — directly inside messaging platforms that support rich message formats (Slack, Discord, Web).
Bearer Token
A type of API authentication credential. Passed in the Authorization: Bearer <token> HTTP header on all API requests.
Bot Token A credential issued by a messaging platform (Telegram, Slack, Discord) that allows OpenClaw to act as a bot on that platform. Never share your bot tokens publicly.
Canvas See Live Canvas.
Channel
A messaging platform connected to OpenClaw. Examples include Telegram, WhatsApp, Slack, Discord, iMessage, Signal, Microsoft Teams, and the built-in Web UI. Each channel has its own configuration block in openclaw.json.
Claude The AI model family made by Anthropic. The recommended provider for OpenClaw. Available in three tiers — Haiku (fast/cheap), Sonnet (balanced), Opus (powerful/expensive).
ClawHub
The official OpenClaw skill registry. A curated collection of community-built and officially maintained skills that can be installed with a single command. Browse at clawhub.dev.
Config File
The openclaw.json file that defines every aspect of your OpenClaw deployment: agents, channels, workspaces, skills, and more. The single source of truth for your gateway.
Context Window The maximum number of tokens (words, roughly) that an AI model can "see" at once. Longer conversations that exceed the context window get truncated. OpenClaw manages context automatically, summarizing old messages to stay within limits.
Control UI
The built-in browser dashboard for OpenClaw. Accessible at http://localhost:3000/dashboard by default. Lets you chat with your agent, manage channels, view logs, and edit configuration — all without touching the terminal.
Default Workspace
The workspace a message is routed to when no other routing rule matches. Every channel has a defaultWorkspace setting.
Docker
A containerization platform used to run OpenClaw in production. The official OpenClaw Docker image at openclaw/openclaw is the recommended deployment method for servers.
Environment Variable
A value defined outside the config file (in .env or the shell environment) and referenced inside openclaw.json with ${VARIABLE_NAME}. Used to keep secrets out of config files.
Fallback Agent
An agent used when the primary agent fails or is unavailable. Configured with "fallback": "agent-name" in the workspace config.
Gateway The central OpenClaw server. Receives messages from all connected channels, routes them to the appropriate agent, collects the response, and sends it back. The gateway is what you install and host.
Haiku Claude's fastest and cheapest model tier. Best for quick answers, simple tasks, and high-volume workloads where cost matters most.
iMessage Apple's messaging platform. OpenClaw integrates with iMessage via the macOS node feature — your Mac receives iMessages and forwards them to the gateway.
Live Canvas OpenClaw's real-time visual collaboration surface. Allows the agent to draw diagrams, charts, and annotated images that users can view in a browser and collaborate on.
Memory Skill A skill that gives the agent persistent memory across sessions. The agent can remember facts about users, preferences, and ongoing projects — even after a conversation ends.
Model
The specific AI version used by an agent. Examples: claude-sonnet-4-6, gpt-4o, gemini-1.5-pro, llama3.1:8b (local via Ollama).
Multi-Agent Routing The ability to route messages to different agents based on rules — workspace, channel, message content, or time of day.
Node An OpenClaw instance running on a device. Usually refers to the main gateway, but can also mean a mobile node (iOS/Android) or a secondary instance.
Ollama A tool for running AI models locally on your own hardware. OpenClaw integrates with Ollama to use local models (Llama, Mistral, Phi, etc.) without sending data to cloud providers.
Opus Claude's most powerful model tier. Best for complex reasoning, long documents, code generation, and tasks requiring deep understanding. More expensive than Haiku and Sonnet.
Pairing The process of authorizing a new messaging account to interact with OpenClaw. A pairing code is generated on the gateway side and entered by the user in the chat. After pairing, the user's ID is added to the allowlist.
Persistent Memory See Memory Skill.
Provider See Agent Provider.
Rate Limiting A mechanism that caps the number of messages a user or workspace can send per minute/hour. Prevents abuse and controls costs. Configured in the workspace config.
Redis An in-memory data store used by OpenClaw for session storage, caching, and real-time features (canvas, collaboration). Required for multi-instance deployments.
Routing The process of deciding which workspace and agent handles an incoming message. Based on rules defined in the workspace config.
Sandbox Mode A safety mode that restricts what the agent can do. In sandbox mode, the agent cannot execute shell commands, write files, or access the network. Useful for public-facing deployments.
Session A conversation context. OpenClaw maintains session state per user, per workspace. The agent remembers the conversation history within a session, giving it context for each new message.
Session Timeout How long a session stays active after the last message. After the timeout, the session is cleared and the next message starts a fresh conversation.
Signal An encrypted messaging app. OpenClaw can connect to Signal via the Signal CLI bridge.
Skill
A modular capability that extends what the agent can do. Examples: web-search, github, bash, memory, canvas, image-generation. Skills are enabled per workspace. Browse skills at ClawHub.
Slash Command
A message that starts with /. OpenClaw recognizes built-in slash commands (like /help, /clear, /canvas new) and routes them to specific handlers instead of the AI agent.
Sonnet Claude's balanced model tier. The default for most OpenClaw deployments. Good balance of capability, speed, and cost.
System Prompt An instruction given to the agent before any user message. Defines the agent's persona, capabilities, and constraints. Set per workspace.
Tailscale A zero-config VPN built on WireGuard. Used to access your OpenClaw Control UI and gateway from anywhere without exposing ports to the public internet.
Token The unit of text that AI models process. Roughly 4 characters or 0.75 words. Models have context window limits measured in tokens, and billing is also per token.
Turbopack
A fast JavaScript bundler used by Next.js. Note: Turbopack requires native binaries and may not work on all platforms — use --webpack flag if you encounter issues.
Voice Mode OpenClaw's feature for interacting via spoken language. On macOS, you can speak to OpenClaw and hear responses. On iOS/Android, voice input goes through the native keyboard or a companion app.
Watchtower A Docker tool that automatically updates containers when new images are available. Can be used to keep OpenClaw up to date automatically.
Webhook An HTTP callback triggered by an event. OpenClaw supports outbound webhooks (notifying your app when events happen) and inbound webhooks (receiving events from external services).
Workspace A logical grouping in OpenClaw that defines: which users can access it (allowlist), which agent handles messages, which skills are available, what the system prompt is, and rate limits. One gateway can have many workspaces serving different teams or use cases.
Zero-Downtime Deployment
A deployment strategy where the gateway continues serving requests while being updated. Achieved in Docker with docker compose up -d --remove-orphans which replaces containers one at a time.