OpenClaw Academy
Glossary
Every OpenClaw term explained. Bookmark this.
47 terms · Click any term heading to copy its direct link
AGENTS.md
A workspace file that contains operating instructions for the agent — how it should behave, what rules to follow, how to use memory, and what priorities to enforce. Loaded at the start of every session. Think of it as the agent's playbook.
Agent
A fully scoped AI brain with its own workspace, session store, auth profiles, and identity. Each agent is isolated — separate files, separate memory, separate personality. A single Gateway can run one or many agents side-by-side.
Agent Loop
The full lifecycle of a single agent run: intake → context assembly → model inference → tool execution → streaming replies → persistence. Each loop is serialized per session to prevent race conditions and keep history consistent.
Binding
A routing rule that maps inbound messages to a specific agent. Bindings match on channel, account ID, peer (DM or group), guild ID (Discord), or team ID (Slack). Most-specific match wins. This is how one Gateway routes messages to the right agent.
Bootstrap Files
The set of workspace files injected into the agent's context at session start: AGENTS.md, SOUL.md, USER.md, TOOLS.md, IDENTITY.md, and HEARTBEAT.md. They give the model its instructions, persona, and user profile without needing to read files manually.
Broadcast Group
A configuration that lets multiple agents respond to the same chat (e.g. a WhatsApp group). When OpenClaw would normally reply, all agents in the broadcast group run in parallel. Useful for multi-persona setups or logging alongside responses.
Canvas
A web-based UI surface that nodes can display. The Gateway serves canvas files from the workspace's canvas/ directory over HTTP. Agents can push content to node displays using the canvas tool — useful for dashboards, visual feedback, or interactive UIs on companion devices.
Channel
A messaging platform connected to OpenClaw: WhatsApp, Telegram, Discord, Slack, Signal, iMessage, WebChat, Matrix, Nostr, Google Chat, or Microsoft Teams. Each channel has its own connector, formatting rules, and access policies. The agent replies on whatever channel the message came from.
ClawHub
The public skill registry for OpenClaw at clawhub.com. Browse, search, install, update, and publish skills. All skills are public and free. Use the CLI (clawhub search, clawhub install) to add capabilities to your agent.
Compaction
When a session nears the model's context window limit, OpenClaw summarizes older conversation into a compact entry and keeps recent messages intact. The summary persists in the JSONL history. You can also trigger it manually with /compact.
Context Window
The maximum number of tokens a model can see in a single run. Everything — system prompt, conversation history, tool results, and attachments — must fit inside this window. Use /status to check usage and /compact to free space.
Cron
The Gateway's built-in job scheduler. Cron jobs persist to disk and survive restarts. Use them for precise timing — daily reports, one-shot reminders, or recurring tasks. Jobs can run in the main session or in isolated sessions, and optionally deliver output to a chat.
DM Pairing
An owner-approval step for inbound DMs. When a channel uses the 'pairing' DM policy, unknown senders receive a short code and their message isn't processed until the owner approves. Pairing codes expire after 1 hour. This is how you control who can talk to your bot.
DM Policy
Controls who can send direct messages to the agent on a given channel. Options include open (anyone), allowlist (pre-approved senders only), and pairing (approval-on-first-contact). Set per-channel in configuration.
Docker Sandbox
An optional Docker container that isolates tool execution from the host. When enabled, exec, read, write, and edit all run inside the container. Configurable per mode (off/non-main/all), scope (session/agent/shared), and workspace access (none/ro/rw). Not a perfect security boundary, but materially limits blast radius.
Elevated Mode
A special execution mode that lets the agent run commands directly on the Gateway host, bypassing the sandbox. Gated by tools.elevated allowlists and sender identity. Use sparingly — elevated exec has full host access.
Exec
The shell execution tool. Runs commands in the workspace (or sandbox). Supports foreground and background modes, PTY for interactive CLIs, environment overrides, timeouts, and multi-host execution (sandbox, gateway, or node). The most powerful — and most dangerous — tool in the kit.
Gateway
The always-on process at the heart of OpenClaw. It owns messaging connections (WhatsApp, Telegram, etc.), the agent runtime, session management, cron scheduling, node coordination, and the WebSocket control plane. Everything flows through the Gateway. Start it with openclaw gateway.
Heartbeat
A periodic agent run (default: every 30 minutes) in the main session. The agent checks a HEARTBEAT.md checklist — inbox, calendar, notifications — and reports anything important. If nothing needs attention, it replies HEARTBEAT_OK silently. Cheaper and more context-aware than multiple cron jobs.
IDENTITY.md
A workspace file that stores the agent's name, vibe, and emoji. Created during the bootstrap ritual and loaded every session. Keeps the agent's identity consistent across sessions.
MCP (Model Context Protocol)
A standardized protocol for connecting AI models to external tools and data sources. OpenClaw supports MCP servers, allowing agents to access databases, APIs, and services through a unified interface. MCP extends the agent's capabilities beyond built-in tools.
Memory
Plain Markdown files in the workspace that persist knowledge across sessions. Daily logs go in memory/YYYY-MM-DD.md; curated long-term memory goes in MEMORY.md. The model only 'remembers' what's written to disk. OpenClaw also supports vector-based semantic search over memory files.
Memory Flush
An automatic silent agent turn triggered when a session nears compaction. The model is reminded to write durable notes to disk before context is summarized. Runs once per compaction cycle and stays invisible to the user.
Memory Search
Semantic search over workspace memory files using vector embeddings. Supports hybrid search (BM25 keyword + vector similarity), local or remote embeddings, and optional session transcript indexing. Tools: memory_search and memory_get.
Model Provider
An LLM service that powers the agent's reasoning. OpenClaw supports OpenAI, Anthropic, Google (Gemini), OpenRouter, AWS Bedrock, Azure, GitHub Copilot, and custom OpenAI-compatible endpoints. Model refs use provider/model format (e.g. anthropic/claude-sonnet-4-5).
Multi-Agent
Running multiple isolated agents in one Gateway. Each agent gets its own workspace, sessions, auth, and personality. Route messages to agents via bindings. Use cases: different phone numbers for different agents, different personas per channel, or multiple people sharing one server.
Node
A companion device (iOS, Android, macOS, or headless) that connects to the Gateway via WebSocket. Nodes expose commands like camera capture, screen recording, location, canvas display, and remote shell execution. They're peripherals — they extend the agent's senses, not its brain.
Onboarding
The guided setup flow for new OpenClaw installations. Run openclaw setup or openclaw onboard to create configuration, seed workspace files, configure model providers, and connect messaging channels. The BOOTSTRAP.md file triggers a one-time first-run ritual.
Orchestrator
An agent pattern where a main agent delegates tasks to specialist sub-agents instead of doing everything itself. The orchestrator stays responsive and available while sub-agents handle research, coding, design, and other work in parallel. The 'chief of staff' model.
Pi Agent (pi-mono)
The embedded agent runtime that OpenClaw is built on, derived from the pi-mono codebase. OpenClaw reuses the model and tool infrastructure but owns session management, discovery, and tool wiring. Legacy Pi/Tau settings are not consulted.
Plugin
A TypeScript module that extends OpenClaw at runtime. Plugins can add Gateway RPC methods, chat commands, tool hooks, channel connectors, and skills. Official plugins include Voice Call, Matrix, Nostr, Microsoft Teams, and memory backends. Install with openclaw plugins install.
Prompt Injection
An attack where malicious input tricks the AI model into ignoring its instructions and following the attacker's commands instead. In OpenClaw's context, this can mean an untrusted message in a group chat or a crafted webpage triggering tool calls. Mitigate with sandboxing, tool policy, DM policy, and group allowlists.
Sandbox
The isolation layer between the agent's tool execution and the host system. Can be Docker-based (recommended) or a workspace-scoped restriction. Modes: off, non-main (only non-DM sessions), or all. Scope: per-session, per-agent, or shared. Off by default — you must enable it.
Security
OpenClaw's layered defense model. Run openclaw security audit to check your posture. Key layers: DM policy (who talks), group policy (which groups), tool policy (what tools), sandboxing (where code runs), elevated allowlists (who gets host access), and network exposure (what's reachable).
Session
A conversation context bucket. DMs collapse to the agent's main session for continuity. Groups and channels get their own sessions. Sessions store transcripts as JSONL, track token usage, and reset daily (4 AM local) or on idle timeout. Reset with /new or /reset.
Skill
A folder containing a SKILL.md file (with YAML frontmatter) that teaches the agent how to use a specific tool or workflow. Skills are loaded from three locations: bundled (shipped), managed (~/.openclaw/skills), and workspace (workspace/skills). Workspace skills win on name conflicts. Compatible with the AgentSkills format.
Slash Commands
Chat commands prefixed with / that control the agent and Gateway. Key commands: /status, /new, /compact, /context, /stop, /subagents, /usage, /tts, /help. Some platforms (Discord, Telegram) also support native slash commands.
SOUL.md
A workspace file that defines the agent's persona, tone, and boundaries. Loaded every session. This is where you set personality, communication style, and hard limits on behavior. The agent's soul, literally.
Sub-Agent
A background agent run spawned from an existing session using sessions_spawn. Sub-agents run in isolated sessions, can use a different (cheaper) model, and announce results back to the requester chat when done. They cannot spawn further sub-agents (no recursive fan-out).
System Prompt
The instruction block OpenClaw builds for every agent run. Includes tooling descriptions, safety guardrails, skills list, workspace info, sandbox status, current time, runtime details, and injected bootstrap files. The model sees this as its foundational context. Inspect with /context detail.
Tool
A capability the agent can invoke during a run. Core tools: read, write, edit, exec, process, browser, canvas, nodes, message, web_search, web_fetch, image, tts. Skills teach the agent how to use tools; tool policy controls which tools are available.
Tool Policy
Configuration that controls which tools an agent can use. Set via tools.allow (whitelist) and tools.deny (blacklist) at the global or per-agent level. TOOLS.md is user guidance only — it does not enforce policy. For hard enforcement, use config, sandboxing, and exec approvals.
TOOLS.md
A workspace file for user-maintained notes about local tools, conventions, and integrations. Loaded at session start as guidance. Important: TOOLS.md does NOT control tool availability — it's just notes for the model. Use tool policy config for enforcement.
USER.md
A workspace file that tells the agent who the user is and how to address them. Loaded every session. Include name, preferences, timezone, and any context that helps the agent personalize responses.
Webhook
An HTTP endpoint exposed by the Gateway that lets external systems trigger agent runs. Requires a shared secret token for auth. Use webhooks to connect CI/CD pipelines, monitoring alerts, form submissions, or any external event to your agent.
Workspace
The agent's home directory. Contains bootstrap files (AGENTS.md, SOUL.md, etc.), memory files, skills, and canvas assets. It's the default working directory for all tool operations. Each agent has its own workspace. Default location: ~/.openclaw/workspace. Back it up with a private git repo.
Disclaimer: OpenClaw Academy is a community project, not officially affiliated with OpenClaw. This glossary is community-maintained and may not reflect the latest official OpenClaw terminology.
Missing a term?
OpenClaw moves fast. If you think we're missing something, let us know on Discord.