Skip to content

what-is-clawdbot.md

Latest commit

 

History

History
216 lines (162 loc) · 8.73 KB

what-is-clawdbot.md

File metadata and controls

216 lines (162 loc) · 8.73 KB
Raw

What is OpenClaw? (plain English, beginner-first)

Table of contents (Explain OpenClaw)

OpenClaw is a framework for running a personal AI assistant on hardware you control.

If you've only used ChatGPT/Claude/Gemini in a browser:

  • those products are apps hosted by someone else
  • OpenClaw is the platform you run yourself, and you plug your chosen model/provider into it

OpenClaw's superpower is not "a smarter model." It's that it can:

  • live where you already talk (WhatsApp, Telegram, Discord, iMessage, …)
  • stay always-on (Gateway service — can run as a system daemon via openclaw gateway install on systemd/launchd/schtasks for automatic start on boot)
  • keep state (sessions, memory, policies)
  • optionally take actions (tools, device nodes)

Official overview (good high-level read): https://docs.openclaw.ai

The one-sentence mental model

OpenClaw is a self-hosted Gateway that connects chat apps to an agent that can reason and (optionally) act.

The cast of characters (with analogies)

1) The Gateway = “your assistant’s home base”

The Gateway is a long-running process that you run on a machine you control.

It:

  • receives inbound messages from chat platforms (channels)
  • decides what should happen (routing + policy)
  • runs the agent turn (calls your model provider)
  • sends a response back to the same chat platform
  • stores local state (config, credentials, session logs)

Think of it as:

  • a router + scheduler + policy engine for your assistant

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

2) Channels = "phone lines" into the assistant

A channel is how OpenClaw connects to a messaging platform.

Examples:

  • WhatsApp (via WhatsApp Web / Baileys)
  • Telegram (Bot API)
  • Discord
  • iMessage (macOS integration)
  • Slack
  • Signal
  • Feishu (with streaming cards, reasoning stream, interactive approval/quick-action cards, ACP session binding)
  • Microsoft Teams (v2026.3.24+: official SDK with streaming 1:1 replies, welcome cards, AI labeling, message edit/delete)
  • Webchat (browser-based, built into the Control UI)
  • plus plugins for many more (Matrix with official matrix-js-sdk, Zalo, IRC, Nostr, Twitch, etc.)

Channels normalize “incoming message events” into a common internal shape.

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

3) Sessions = “conversation memory (by default) on disk”

A session is a conversation thread with state:

  • chat history (transcripts)
  • metadata (who/where it came from)
  • optional semantic memory (MEMORY.md + memory/*.md files indexed via embeddings for hybrid keyword + vector search — the agent can recall past context across sessions)

Sessions live on disk under your state directory (usually ~/.openclaw/).

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

4) The agent = "the brain (model + rules + tool policy)"

The agent is where your AI model (Claude/GPT/etc.) is actually called.

OpenClaw supplies:

  • system prompt templates
  • history/context
  • safety wrappers
  • tool availability rules

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

5) Tools = “hands” (powerful; risky)

Tools let the model do more than output text.

Depending on what you enable, tools can include:

  • web fetch/search (including Brave, Exa, Tavily, Firecrawl as bundled web-search plugins; Brave LLM Context mode for grounding snippets)
  • browser automation
  • cron/automation
  • exec or node/device invocations
  • memory search (semantic search across indexed notes and past conversations; supports multimodal image/audio indexing with Gemini embeddings)
  • canvas/drawing (interactive visual outputs)
  • image processing (resize, optimize, convert)
  • PDF analysis (native Anthropic/Google support, extraction fallback for other models)
  • diffs (read-only diff rendering from before/after text or unified patches, with PNG/PDF output)
  • text-to-speech (ElevenLabs, OpenAI, Edge TTS)
  • session management (list, send, history across sessions, yield control to another session)
  • gateway control (config, status, restart from within agent)
  • /btw side questions (ephemeral one-shot queries against current session context, no transcript persistence)

Tools are where most real-world risk comes from.

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

6) Nodes/devices = “peripherals”

Nodes are devices (macOS/iOS/Android) that can connect to the Gateway and offer device-local capabilities.

Examples:

  • camera (snap, clip, list cameras)
  • screen recording
  • location services
  • notifications
  • canvas/webviews
  • (on macOS) remote execution with approvals
  • (on Android, v2026.3.22+) call log search, SMS search, gateway-backed TTS

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

What OpenClaw is great for

Personal / small-team assistant in chat

  • “Summarize the last 200 messages in this group.”
  • “Draft a reply in my tone.”
  • “Keep track of ongoing tasks.”

A private assistant with your own policies

Because you control the Gateway, you can enforce:

  • who can talk to it
  • what it can do
  • where it can reach

An always-on agent host

A VPS or home server can keep the Gateway running even if your laptop sleeps.

What OpenClaw is not

Not automatically safe for public bots

If you open inbound DMs/groups to the public and enable tools, you've basically built a remote-controlled automation engine.

OpenClaw ships safety features (pairing, allowlists, audits), but you must use them.

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

Not "privacy magic"

OpenClaw keeps state locally, but your chosen model provider still receives the prompts you send for inference unless you run a local model.

Recent notable changes (v2026.3.22–3.28-beta.1)

Breaking changes

  • Plugin SDK: openclaw/extension-api removed; use openclaw/plugin-sdk/* subpaths. See migration guide: https://docs.openclaw.ai/plugins/sdk-migration
  • ClawHub-first installs: openclaw plugins install <name> now prefers ClawHub over npm
  • Chrome MCP extension removed: driver: "extension" and browser.relayBindHost gone. Run openclaw doctor --fix to migrate
  • Legacy env vars removed: CLAWDBOT_* and MOLTBOT_* no longer work; use OPENCLAW_*
  • Legacy state dir removed: ~/.moltbot auto-detection removed; use ~/.openclaw or set OPENCLAW_STATE_DIR
  • nano-banana-pro skill removed: use agents.defaults.imageGenerationModel instead
  • Matrix plugin rewritten: now backed by official matrix-js-sdk (migration guide: https://docs.openclaw.ai/install/migrating-matrix)

New capabilities

  • ACP: Spawn sub-agent workspaces in Discord, iMessage, BlueBubbles, Feishu with /acp spawn --bind here
  • Sandbox backends: Pluggable (Docker, OpenShell with mirror/remote, SSH)
  • New providers: Anthropic Vertex AI, Chutes, Xiaomi MiMo V2, gpt-5.4/mini/nano
  • Web search plugins: Exa, Tavily, Firecrawl as bundled providers
  • Microsoft Teams: Official SDK with streaming, welcome cards, AI labeling, edit/delete
  • Telegram: Custom Bot API endpoints (apiRoot), auto topic labels, silent error replies, topic editing
  • Feishu: Interactive approval cards, ACP binding, streaming reasoning tokens
  • Android: Call log search, SMS search, gateway TTS, system dark theme
  • Control UI: Skills management with status tabs, markdown preview, roundness slider, expand-to-canvas
  • MiniMax: Image generation (image-01), M2.7 default model
  • Gateway: OpenAI-compatible /v1/models and /v1/embeddings endpoints
  • CLI: --container for Docker/Podman, config schema, expanded config set

Beginner-safe starting posture (recommended)

If you want high privacy and low risk:

  1. Keep the Gateway loopback-only (localhost only)
  2. Use pairing and/or allowlists so only you can trigger it
  3. Avoid enabling powerful tools until you are comfortable
  4. Run openclaw security audit --deep and fix findings

See:

Next: Glossary | CLI commands