Skip to main content

Glossary

Short definitions of OpenClaw terms, with links to the full documentation. Star counts are as of 2026-10-18.

activeHours​

The window in which the Heartbeat may run, set with agents.defaults.heartbeat.activeHours: { start, end } and checked in the configured timezone. This replaces the older "quiet hours" idea. See Heartbeat.

AMOS​

Atomic macOS Stealer, an infostealer that targets Keychain passwords, browser credentials and crypto wallets. Spread by the ClawHavoc campaign on ClawHub. See Known Vulnerabilities.

Automations​

Scheduled and event-triggered agent jobs, formerly called cron jobs (renamed in v2026.8.1; openclaw cron still works). Recurring checks belong here, not in the heartbeat. See Automations.

Brain​

This site's name for the LLM side of the agent: the model that reasons and decides which tools to call. Upstream calls it the agent runtime and model. Set it with agents.defaults.model (for example anthropic/claude-sonnet-5-5). See Brain & Hands.

Channel​

A connection between OpenClaw and a messaging platform such as WhatsApp, Telegram, Discord or Slack. Configured under channels.<id>; many channels ship as plugins. Email is an automation trigger, not a channel. See Channels & Integrations.

ClawBands​

Community middleware (SeyZ/clawbands, 196 stars) that asks for human approval before dangerous tool calls. No commits since 2026-02. Upstream exec approvals now cover most of this.

Clawdbot​

The original name of OpenClaw (November 2025). Renamed to Moltbot on January 27, 2026 after Anthropic asked for a name change.

Clawdex​

Koi Security's pre-install skill check against its database of known malicious skills. See Skill Verification.

ClawHub​

OpenClaw's skill and plugin registry at clawhub.ai. Since 2026-10-13 a new release stays unpublished until its TruffleHog and ClawScan checks pass; openclaw skills verify @owner/<slug> fetches a release's audit verdict. Target of the ClawHavoc campaign. See ClawHub Guide.

ClawHavoc​

A supply-chain campaign reported by Koi Security in February 2026: 341 malicious skills on ClawHub, 335 of which delivered AMOS through fake "Prerequisites" steps. See Known Vulnerabilities.

ClawPhone​

Community project (marshallrichards/ClawPhone, 549 stars) that runs OpenClaw on a cheap Android phone via Termux. For a supported phone client, use the official Android app (see Companion apps).

Clawprint​

A tamper-evident audit-trail tool, now archived (tsyrulb/clawprint, formerly cyntrisec/clawprint). Historical only; for audit logging use upstream logging.audit and openclaw audit.

ClawSec​

Security skill suite by Prompt Security (prompt-security/clawsec, 1.1k stars): advisory monitoring, audits and SOUL.md drift detection.

Clawery​

clawery.com, an enterprise AI assistant service. Its site no longer mentions OpenClaw. Not part of the open-source project.

Code Insight​

VirusTotal's AI-based package analysis, which early-2026 reports said ClawHub used on uploads. Current upstream docs name VirusTotal, ClawScan and static analysis as ClawHub's scan sources. See Skill Verification.

Companion apps​

Official native clients: the macOS app (macOS 15.0+ in v2026.10.4; the 2026.10.10 beta raises it to 26.2), the iOS app on the App Store, the Android app on Google Play, and the Windows Hub. See Platforms.

openclaw.json​

The main config file, ~/.openclaw/openclaw.json (JSON5). Holds models, channels, gateway, tools and heartbeat settings (for example agents.defaults.heartbeat.every). Override the path with OPENCLAW_CONFIG_PATH; check it with openclaw config validate. See Configuration Reference.

Context Accumulation​

Session history growing over a long conversation, which raises the token count of every call. A main cost driver; keep sessions short or /compact them. See Cost Management.

Control Plane​

The Gateway's WebSocket endpoint (default ws://127.0.0.1:18789). The Control UI, CLI, apps and nodes all connect to it. See Gateway.

Control UI​

The browser app served by the Gateway at http://127.0.0.1:18789/: chat, approvals, settings, files and dashboards. Rebuilt in v2026.8.1. See Control UI.

CVE-2026-25253​

High-severity (CVSS 8.8) one-click RCE fixed in v2026.1.29 on January 30, 2026. The Control UI trusted a gatewayUrl query parameter and sent the Gateway token to it, so a crafted link could steal the token, even from loopback-only instances. See Known Vulnerabilities.

Daemon Mode​

Running the Gateway as a background service (launchd, systemd or Scheduled Tasks) that starts on login or boot. Install it with openclaw gateway install (or the equivalent openclaw daemon install); manage it with openclaw gateway status|stop|start|restart. There is no --daemon flag. See openclaw gateway.

Doctor​

openclaw doctor checks config, services and state. openclaw doctor --fix applies repairs and migrations, such as retired config keys, agents.list, and HEARTBEAT.md. Run it after skipping releases. See Doctor.

Exec approvals​

The policy that decides whether the agent may run a shell command on the host. tools.exec.mode is deny, allowlist, ask, auto or full; inspect it with openclaw approvals get. See Exec approvals and hardening.

Gateway​

The single long-running Node.js process at the centre of OpenClaw. It runs the agent, owns channels, sessions and tools, and exposes the control plane on port 18789. Binds to loopback by default (gateway.bind). See Gateway Architecture.

Hands​

This site's name for the agent's tools: shell (exec), file read/write, browser, web_fetch, web_search and so on. What they may do is set by exec approvals, permission modes and the sandbox. See Brain & Hands.

HEARTBEAT.md​

Retired. A workspace file that used to hold the heartbeat checklist. Since v2026.8.1 the runtime never reads it. openclaw doctor --fix imports it into the heartbeat's monitor scratch, turns tasks: entries into automations, and archives the file. See Heartbeat.

Heartbeat​

A periodic agent turn that checks whether anything needs attention. Configured with agents.defaults.heartbeat (every, default 30m; "0m" turns it off; plus activeHours, model, isolatedSession, lightContext). When nothing needs doing the agent replies NO_REPLY. A notable cost driver. See Heartbeat and Heartbeat Architecture.

Lobster​

OpenClaw's official workflow shell (openclaw/lobster, 1.3k stars): a typed, local-first macro engine that chains skills and tools into pipelines. See Lobster Workflows.

Memory​

Long-term notes in the agent workspace: ~/.openclaw/workspace/MEMORY.md plus daily files in memory/YYYY-MM-DD.md, searched by the built-in memory. Memory is stored locally but is injected into prompts, so it reaches your model provider. See Memory System and Agent workspace.

memU​

Third-party memory framework for proactive agents (NevaMind-AI/memU, 14.6k stars). Check its LICENSE file before use.

MimiClaw​

A separate agent written in C for the ESP32-S3 microcontroller (memovai/mimiclaw, 5.8k stars). Not OpenClaw itself. See Lightweight Variants.

Moltbook​

A separate social network (created by Matt Schlicht) where AI agents post, comment and vote. Suffered a major database breach in January 2026.

Moltbot​

OpenClaw's second name (January 27-30, 2026), adopted after the Clawdbot trademark request, then dropped because "Moltbot never quite rolled off the tongue".

Moltworker​

cloudflare/moltworker (9.9k stars) runs OpenClaw in a Cloudflare Sandbox on Workers. Its README calls it an experimental proof of concept, not officially supported. About $34.50/month running 24/7 per the README. Last pushed 2026-05.

Monitor scratch​

The heartbeat's instruction text, stored in the state database instead of a file. The heartbeat runs as an automation job: find its id with openclaw cron list --all, then read the scratch with openclaw cron scratch <jobId> or replace it with --set "...". Replaces HEARTBEAT.md. See Heartbeat.

NanoClaw​

An independent lightweight agent that runs each agent in a container (nanocoai/nanoclaw, formerly gavrielc/nanoclaw, 31k stars). See Lightweight Variants.

Nanobot​

An independent lightweight Python agent from the University of Hong Kong (HKUDS/nanobot, 49k stars). See Lightweight Variants.

NO_REPLY​

The reply a heartbeat turn gives when nothing needs attention; OpenClaw then sends nothing. HEARTBEAT_OK is the legacy form.

OpenClaw​

Free, open-source AI agent you run yourself, connecting LLMs to your files, shell, browser and messaging apps. About 392k GitHub stars as of 2026-10-18. Created by Peter Steinberger, stewarded by the OpenClaw Foundation, MIT licensed. Previously Clawdbot and Moltbot.

Orchestrator​

This site's name for the part of the Gateway that routes work between the model, tools and memory. See Architecture Overview.

Permission mode​

A per-session access level: read-only, guarded, workspace or full, setting the session's filesystem boundary and who reviews exec escalations. Separate from tools.exec.mode, which governs host exec approvals. See Session permission modes.

Profile​

A named, isolated OpenClaw instance with its own config, state and workspace (~/.openclaw-<name>/). Select it with openclaw --profile <name> … or OPENCLAW_PROFILE. Used to run several gateways on one machine. See Multiple gateways.

Prompt Injection​

An attack where untrusted content (messages, web pages, skill output, documents) steers the agent into unintended actions. Trusted and untrusted text share one model context. See Security Overview.

Quiet Hours​

Older name for activeHours: the heartbeat runs only inside agents.defaults.heartbeat.activeHours.

Router​

This site's name for the Gateway logic that decides which agent and session handle an incoming message. Upstream configures it with bindings (multi-agent routing). See Architecture Overview.

Sandbox Mode​

Running an agent's tool execution inside a container instead of on the host, set with agents.defaults.sandbox.mode (off by default, non-main or all) or per agent. It applies per agent or session, not per skill; there is no --sandbox install flag. See Security Hardening: sandboxing and Skill Verification.

Skill​

A folder containing a SKILL.md file (YAML frontmatter plus instructions) and any supporting files, which teaches the agent a task. Install from ClawHub with openclaw skills install, or write your own. See Skill Development.

SkillGuard​

A skill scanner (bossondehiggs/skillguard) with 0 stars and a single push in February 2026. Effectively abandoned; don't rely on it. Use openclaw skills verify and openclaw security audit --deep.

SOUL.md​

The agent's identity and tone file, ~/.openclaw/workspace/SOUL.md. A prompt injection target: an attacker who can write to it can change the agent's behaviour persistently. See SOUL.md Guide.

ToxicSkills​

Snyk's February 5, 2026 audit of 3,984 skills from ClawHub and skills.sh: 36% had security flaws, 13.4% critical. See Known Vulnerabilities.

trustedProxies​

gateway.trustedProxies lists the reverse-proxy IPs whose forwarding headers the Gateway trusts. Without it, a proxied gateway can treat every request as local or accept spoofed X-Forwarded-For headers. See Reverse proxies.

VisionClaw​

Community project for Meta Ray-Ban smart glasses with voice and vision (Intent-Lab/VisionClaw, formerly sseanliu/VisionClaw, 2.6k stars).

WebChat​

The chat surface over the Gateway WebSocket, available in the Control UI and the native apps. The old gateway.webchat and channels.webchat config keys are retired. See WebChat.

WebClaw​

Third-party browser client for the Gateway (ibelick/webclaw, 647 stars). No commits since 2026-03. It only speaks Gateway protocol 3, so it cannot connect to current Gateways; use the built-in Control UI. See WebClaw Guide.

Windows Hub​

The official native Windows app (openclaw/openclaw-windows-node, 2.2k stars, formerly shanselman/openclaw-windows-hub). Upstream recommends it for Windows 10 20H2+ and 11. See Windows.

See Also​