Glossary

A

A2A (Agent-to-Agent)

A standardized inter-agent communication protocol based on JSON-RPC 2.0 over SSE, enabling structured conversations between the workspace and third-party AI Agents. → A2A Protocol

Agent Card

The identity description card of an agent in the A2A protocol, containing the avatar, name, description, and capability list. Retrieved via an HTTP endpoint.

Agent Loop

The AI Agent's autonomous reasoning-execution loop mechanism. Within the loop, the Agent analyzes tasks, selects tools, performs operations, and evaluates results, with a default maximum of 25 rounds. → Agent Loop Engine

B

BM25

A classic keyword-matching algorithm used as a fallback for LLM semantic routing in the Gateway's intelligent routing. Parameters: k1=1.5, b=0.75.

C

Circuit Breaker

A protection mechanism that prevents consecutive LLM failures from causing infinite retries. A 60-second cooldown period is triggered after 5 consecutive failures.

Work

The core conversation interface of the workspace, through which users interact with AI Agents. → Work Conversation Overview

claw-shared

The shared UI component layer for the Web and APP clients. It achieves platform-agnostic component reuse through the PlatformAdapter pattern.

Context Window

The maximum number of tokens an LLM can process in a single inference. Context windows vary across models, ranging from 64K to 1M. → Supported Models List

D

Distill

The workspace's knowledge distillation feature module (Coming Soon), planned to provide AI-driven knowledge extraction and summarization.

E

EMA (Exponential Moving Average)

An algorithm used to dynamically calibrate token estimation accuracy. Initial value 3.0 chars/token, α=0.15.

Enterprise Memory

An organization-level memory dimension, isolated by orgId and maintainable only by administrators. → Three-Dimensional Memory System

G

Gateway

The WebSocket gateway service, serving as the central dispatch hub of the multi-node architecture. It is responsible for node registration, intelligent routing, message forwarding, and permission verification. → Multi-Node Architecture

H

Hub-Spoke Topology

The visualization layout method for the memory graph. A Hub (central node) radiates connections to multiple Facts (fact nodes), and Entities (entity nodes) are cross-linked with one another.

I

InlinePreviewCard

A thumbnail display component for files in the conversation stream, supporting hover highlighting and action buttons (expand, copy, show in file manager, etc.).

M

MCP (Model Context Protocol)

A standardized AI tool communication protocol supporting three transport methods (stdio / SSE / HTTP) and five authentication modes. → Tool Management

mem0

The memory storage backend service, providing semantic search, knowledge graph, and automatic update/merge/forget capabilities.

Multimodal

The capability to input and output multiple content types (text, images, documents, audio, video, etc.). → Multimodal Input and Output

N

Node

A device unit in the multi-node architecture. The Web client is a Human-type node, and the APP client is an Agent-type node. → Multi-Node Architecture

P

PlatformAdapter

The platform adaptation pattern for Work shared components. The WorkPage reads state and operations through usePlatformAdapter() without directly depending on platform code.

P0 / P1 / P2

Tool security levels. P0 = automatic execution, P1 = confirm on first use, P2 = confirm every time. → Runtime Security

S

Sandbox

A lightweight micro-VM based on VirtIO, providing an isolated code execution environment and supporting arm64/amd64. → Runtime Security

Sidecar

The Node.js sidecar process on the APP client, running core components such as the Agent Engine, Context Manager, and Tool Handler. Its lifecycle is managed by the Tauri Rust process.

Skill

A pre-packaged professional capability module, declared via a SKILL.md file. The system comes with 21 built-in skills. → Skill Management

Subagent

The Agent's task decomposition and delegation mechanism. It supports three types—automatic creation, pre-registered dispatch, and cross-node dispatch—with a maximum nesting of 3 levels. → Subagent System

T

Tool

A capability unit through which the Agent performs operations. The system has 19 built-in tools, extensible via MCP, API, and skills. → Tool Management

W

Workflow

A visual multi-step automation process that supports orchestrating multiple components and data sources. → Workflow