Glossary

This glossary covers the key terms used across MultiClaw and its documentation. If you hit an unfamiliar word in another article, start here.

Agent

An AI assistant you configure to perform tasks on your behalf. Each agent has a name, a personality (system prompt), an AI model, and a set of tools it can use.

The language model (for example, from OpenAI or Anthropic) that powers an agent's reasoning and actions. You choose a model when you create an agent and can change it later from the agent's settings. Workspace guardrails can restrict which model providers are available. See also: Token.

Approval

Your explicit confirmation that an agent's proposed plan is ready to execute. When an agent presents a plan during a conversation, you can review and confirm before the agent proceeds — or redirect the agent with further instructions.

Audit trail

A time-stamped record of agent actions, tool calls, and conversation events. Useful when you need to review what an agent did and when.

Browser automation

Agents can control a real web browser to complete web-based tasks on your behalf: navigating pages, clicking buttons, and filling forms. When using a cloud desktop, the browser runs inside the remote virtual machine, not on your local machine. See also: Recording, Skill.

Channel

An application or platform an agent can interact with. The currently supported channel is Slack, which lets agents respond to direct messages and channel mentions.

Cloud desktop

A remote virtual machine managed by MultiClaw where agents can run continuously, independent of your local machine. Also called an "instance" in some parts of the interface.

Connected

The state when the desktop app is signed in to MultiClaw Cloud and the OpenClaw gateway is running. Agents can only run tasks when connected.

Conversation

A chat thread between you and an agent. Contains your messages and the agent's responses, including any tool calls the agent makes along the way.

Dashboard

The workspace home view in MultiClaw Cloud. Provides navigation to your workspace's agents, instances, skills, users, usage, and guardrails.

Desktop app

The MultiClaw application installed on your Mac, Windows, or Linux machine. The primary interface for chatting with agents and managing your workspace.

Gateway

The local web server that OpenClaw runs on your machine. The desktop app connects to it to communicate with agents. If the gateway is not running, agents cannot respond.

Guardrails

Rules set by a workspace owner that restrict what agents can do, applying across all workspace members. Examples: limit which AI models agents may use, block specific tools or categories of tools, require human approval for certain action types, or cap the token budget per task.

Instance

See Cloud desktop. Used interchangeably in some parts of the interface.

Local agent

An agent stored and running on your own machine rather than in MultiClaw Cloud. Local agents are only accessible on the device where they are stored, do not sync across devices, and only run while your machine is on. See also: Cloud desktop.

MCP (Model Context Protocol)

An open protocol that lets agents connect to external tool servers. Each MCP server exposes a set of tools agents can call.

MCP server

A server that implements the Model Context Protocol and provides agents with access to specific tools or data sources (for example, GitHub, Google Drive, or a database).

MultiClaw Chrome Extension

A browser extension (Chrome, Edge, Brave, and other Chromium-based browsers) used to record browser actions so you can teach agents new workflows and save them as skills.

MultiClaw Cloud

The cloud platform at multiclaw.ai that manages your account, workspace, team, and cloud desktops.

OpenClaw

The open-source AI agent engine that powers MultiClaw. Runs locally or on a cloud desktop and stores all agent data as plain files. Managed automatically by the desktop app.

Personality

The system prompt that defines how an agent thinks, responds, and behaves. Written in plain English and set from the Personality tab on the agent's profile. A focused, role-specific personality produces more consistent results than a generic one.

Plan

An ordered list of steps an agent proposes within a conversation before carrying out a task. You can review and redirect the plan before the agent acts.

Quick chat

A conversational interaction where the agent answers questions, brainstorms, or summarizes rather than taking action on your behalf. Contrast with a task, where the agent drafts a plan and executes steps. Both happen inside a conversation — the difference is whether you ask a question or tell the agent to do something.

Recording

A captured sequence of browser actions created with the MultiClaw Chrome Extension. You can convert a recording into a reusable skill.

Role

Each workspace member has a role that controls what they can see and do. Two roles exist: Owner and Member. The Owner created the workspace and can invite or remove members. Members can use workspace resources but cannot manage membership.

Skill

A reusable procedure that an agent can follow, created from a browser recording or written manually. Stored as a SKILL.md file in OpenClaw.

Subagent

An agent that another agent can delegate subtasks to. When subagents are configured, the parent agent can break a large task into smaller pieces and propose which parts to hand off. This delegation appears in the plan you approve before execution begins. You can configure subagents from the Subagents tab on the agent's profile.

Task

A unit of work that an agent carries out during a conversation. When you ask an agent to act, the agent drafts a plan outlining the proposed steps. After you review and approve the plan, the agent executes it on your local machine or a cloud desktop, depending on the agent type. Results stream back to your conversation in real time. See also: Plan, Approval.