Skip to main content

📝 OpenClaw

Description​

< What is it? >​

OpenClaw is an open-source, self-hosted AI assistant platform. Its Gateway connects chat surfaces—such as Discord, Slack, Telegram, WhatsApp, or a web dashboard—to an agent runtime. The Gateway owns session routing and channel connections; configured agents can use selected models, tools, skills, plugins, and automation.

It is a complete agent runtime and deployment surface—not an LLM, vector database, or Model Context Protocol (MCP) server. You run it on your own machine or server and decide which channels, credentials, workspaces, and capabilities an agent may use.

Key points​

< Gateway-centered design >​

PartRole
Chat channel or dashboardWhere a person sends a request and receives the response.
GatewayThe central control plane for channel connections, sessions, routing, and configured policy.
Agent runtimeUses an LLM to decide how to respond or which enabled tool to call.
Tools, skills, and pluginsGive the agent bounded capabilities, reusable procedures, and integrations.
Model providerSupplies the model used by the agent; OpenClaw can be configured with different supported providers.
Chat app or dashboard
        ↓
      Gateway
        ↓
 Agent runtime + model
        ↓
Tools, skills, plugins, and permitted services

< Simple examples >​

GoalExample flow
Ask from a chat appYou message: “Summarize the open issues for project Alpha.” The Gateway routes the message to an agent session. If an issue-tracker tool is installed and allowed, the agent retrieves the issues and replies in the same chat.
Use a coding assistant remotelyYou message a configured agent from Telegram: “Explain the failing test in this workspace.” The agent can inspect only the workspace and tools it has been granted, then sends the explanation back through Telegram.
Run a recurring taskA scheduled job asks an agent to prepare a daily briefing from permitted sources, then posts the result to a configured Slack channel.
Add retrieval to an agentAn agent uses a LlamaIndex retriever as a tool, retrieves relevant document chunks, and answers a question using that context.

OpenClaw provides the agent-facing runtime and routing. The individual tools, data sources, and approval rules still need to be configured for the application.

< Security boundaries matter >​

An agent connected to messages, files, accounts, or a shell can affect real systems. Self-hosting gives the operator control, but it does not automatically make an agent safe.

  • Allow only trusted senders and channels.
  • Give each agent the least access it needs: prefer read-only tools and scoped workspaces.
  • Require approvals for irreversible actions such as sending messages, deleting data, spending money, or deploying code.
  • Treat tool output and retrieved text as untrusted data; they can contain prompt-injection attempts.
  • Start with a low-risk environment before connecting personal or production accounts.

Comparison​

TechnologyPrimary roleHow it relates to OpenClaw
OpenClawSelf-hosted gateway and agent platformRoutes conversations, runs agent sessions, and connects configured capabilities.
LlamaIndexFramework for RAG and context-augmented applicationsCan provide retrieval or query tools that an OpenClaw agent uses.
MCPStandard protocol for connecting AI applications to external capabilitiesCan be one way for an OpenClaw deployment to reach external tools or context.
Agent SkillsReusable instructions and supporting resourcesCan teach an agent repeatable workflows without placing every procedure in each prompt.

Video Tutorial​

Reference​