Skip to content

glossary.md

Latest commit

 

History

History
177 lines (122 loc) · 9.43 KB

glossary.md

File metadata and controls

177 lines (122 loc) · 9.43 KB

Glossary (plain English)

Table of contents (Explain OpenClaw)

This glossary is intentionally biased toward the words you’ll see in the CLI output and docs.

ACP (Agent Communication Protocol)

A protocol for spawning and managing sub-agent sessions within channels. Allows binding an ACP session to the current conversation (e.g., /acp spawn codex --bind here) so a Codex-backed or Claude-backed workspace runs directly in the active Discord thread, iMessage chat, BlueBubbles DM, or Feishu topic — without creating a child thread. Distinguishes between chat surface, ACP session, and runtime workspace.

Docs: https://docs.openclaw.ai/concepts/acp

Agent

The “brain” that turns an incoming message + context into a response. In practice it’s:

  • your chosen model (provider + model name)
  • a system prompt + templates
  • tool availability rules
  • session/memory behavior

Docs: https://docs.openclaw.ai/concepts/agent

Allowlist

A list of identities allowed to trigger the bot (DMs) or allowed to interact in groups.

Security note: allowlists are often the real security boundary in messaging bots.

Docs: https://docs.openclaw.ai/gateway/security

Bind (gateway.bind)

Where the Gateway listens:

  • loopback = localhost only (safest default)
  • lan = your LAN interfaces (requires auth)
  • tailnet = bind only to Tailscale IP, 100.x.y.z range (requires Tailscale)
  • auto = automatically choose the best bind mode based on environment
  • custom = bind to a specific IP address (set via gateway.customBindHost); if unavailable, falls back to 0.0.0.0

Docs: https://docs.openclaw.ai/gateway/remote and https://docs.openclaw.ai/gateway/tailscale

Bundle

A packaged collection of skills, MCP servers, and configuration defaults from compatible AI coding tools (Codex, Claude, Cursor). OpenClaw can discover, install, and map bundle skills into its own skill surface, and bundle MCP servers expose runnable tools in embedded Pi.

Channel

A connector for a messaging surface: WhatsApp/Telegram/Discord/iMessage/etc.

Docs: https://docs.openclaw.ai/channels

ClawHub

The official OpenClaw marketplace for skills and plugins. Since v2026.3.22, openclaw plugins install <package> prefers ClawHub before npm by default. Native commands: openclaw skills search|install|update. Also accessible via openclaw plugins install clawhub:<package> for explicit source targeting.

Docs: https://docs.openclaw.ai/tools/clawhub

Control UI / Dashboard

The web interface served by the Gateway (on the same port as the WebSocket). Dashboard-v2 (v2026.3.12) added modular overview, chat, config, agent, and session views, a command palette, mobile bottom tabs, and richer chat tools (slash commands, search, export, pinned messages).

Docs: https://docs.openclaw.ai/web/dashboard

Fast mode

A per-session toggle (/fast) that requests faster output from supported providers (Anthropic service_tier, OpenAI fast tier). Configurable via TUI, Control UI, ACP, and per-model config defaults.

SecretRef

A reference-based credential system that resolves secrets at runtime from environment variables, files, or exec commands — instead of storing plaintext values in config. Supports 64 credential targets across providers, channels, webhooks, and gateway auth.

Gateway

The long-running process that owns:

  • channel connections
  • routing
  • sessions
  • plugins
  • tool execution policy
  • node/device pairing

Docs: https://docs.openclaw.ai/gateway

Hook

A webhook endpoint on the Gateway that runs automation on inbound HTTP requests. Configure via hooks.mappings to route external events (CI alerts, calendar triggers, IoT) into the agent.

Docs: https://docs.openclaw.ai/gateway/hooks

Model

A specific AI model identified by provider/model-id (e.g., anthropic/claude-sonnet-4-20250514). Set the default via agents.defaults.model.primary, or pass --model per invocation.

Docs: https://docs.openclaw.ai/concepts/model

Pairing

Explicit owner approval.

Used for:

  • DM pairing (who may DM the bot)
  • device pairing (which nodes may connect)

Docs: https://docs.openclaw.ai/start/pairing

Plugin / extension

Extra code that runs in-process with the Gateway to add channels/tools/features.

Treat it like installing code into your assistant.

The public plugin SDK surface is openclaw/plugin-sdk/* (since v2026.3.22). The legacy openclaw/extension-api import path has been removed — plugins must use narrow openclaw/plugin-sdk/* subpaths. Plugin hooks include before_tool_call (with async requireApproval for user confirmation), before_dispatch (with canonical inbound metadata), and standard lifecycle hooks.

Docs: https://docs.openclaw.ai/plugins/sdk-overview and https://docs.openclaw.ai/plugins/sdk-migration

Profile

A named configuration variant selected via the OPENCLAW_PROFILE=<name> environment variable. Each profile stores its config alongside the default, so you can switch between (for example) a local-model profile and a cloud-model profile without editing files.

Provider

An AI service configured under models.providers.<name> (e.g., Anthropic, OpenAI, Ollama, Docker Model Runner). Each provider entry sets baseUrl, apiKey, and the model list it exposes. As of v2026.3.22, several providers are implemented as bundled plugins: OpenRouter, GitHub Copilot, OpenAI Codex, Chutes, MiniMax, and DeepSeek. New providers added in v2026.3.22+: Anthropic Vertex (Claude via Google Vertex AI), Chutes, Xiaomi MiMo V2. Web search providers added as bundled plugins: Exa, Tavily, Firecrawl.

Sandbox

An isolated execution environment for tool runs and subagents with tighter capability limits. Configure via tools.sandbox to give sandboxed sessions a different (usually more restrictive) tool policy than main sessions.

Since v2026.3.22, sandbox backends are pluggable:

  • Docker — the original container-based backend (still supported)
  • OpenShell — newer backend with mirror (sync workspace into sandbox) and remote (use an existing remote shell) workspace modes
  • SSH — core SSH backend with secret-backed key, certificate, and known_hosts inputs for remote sandbox execution

openclaw sandbox list/recreate/prune commands are now backend-aware, not Docker-only.

Docs: https://docs.openclaw.ai/gateway/sandboxing

Session

A stored conversation thread (history + metadata). By default sessions are stored on disk as JSONL.

Docs: https://docs.openclaw.ai/concepts/session

Tool

A capability the model can invoke (web fetch/search, browser automation, cron, exec, node calls).

Docs: https://docs.openclaw.ai/tools

Webchat

A browser-based channel that connects to the Gateway's WebSocket endpoint directly. Unlike external messaging platforms, webchat runs in the Control UI and supports toggles for hiding tool calls and thinking blocks.

Extended thinking / reasoning

When a model supports "extended thinking" (e.g., Claude with thinking or providers with reasoning tokens), OpenClaw can stream intermediate reasoning steps before the final answer. Configured via thinking level settings (--thinking flag or agents.defaults.thinking).

Skill

A packaged AI routine (commands + optional tools) that extends the agent's capabilities. Install from ClawHub or load locally; configured via skills.load.*. The built-in skill scanner checks for dangerous code patterns at install time.

Docs: https://docs.openclaw.ai/skills

Node / device

A companion device connected to the Gateway (iOS/Android/macOS/headless nodes). Android nodes gained callLog.search (call history), sms.search (text messages), and gateway-backed talk.speak (TTS synthesis) in v2026.3.22.

Docs: https://docs.openclaw.ai/nodes

Context engine (plugin)

A plugin that owns how conversation context is assembled before model calls. Plugins can implement delegateCompactionToRuntime(...) for runtime-managed compaction, transcript maintenance rewrites, and model-aware assemble() that receives the modelId to adapt formatting per model.

Health monitor

Per-channel and per-account health monitoring with configurable stale-event thresholds, restart limits, and healthMonitor.enabled overrides. Global disable: gateway.channelHealthCheckMinutes=0.

Image generation

Standardized on the core image_generate tool. Configure via agents.defaults.imageGenerationModel.primary. MiniMax added image-01 model with generate and image-to-image editing. The legacy nano-banana-pro bundled skill was removed in v2026.3.22.

Memory plugin

The memory subsystem is plugin-driven. The active memory plugin (memory-core by default, or memory-lancedb for vector search) owns flush prompts, target-path policy, system-prompt sections, and embedding provider registrations. Configure via plugins.slots.memory.