Tools, skills, plugins, MCP servers — the terminology in the autonomous agent space is confusing. This guide cuts through the noise and explains exactly how OpenClaw's architecture works, what each layer does, and how they fit together. Essential reading for anyone building on, evaluating, or comparing OpenClaw.
OpenClaw's architecture has five distinct layers, and understanding each one is key to using the platform effectively — and to evaluating alternatives like NanoClaw, Moltworker, and Nanobot. Layer 1: The Model (LLM) The AI model that powers reasoning: GPT-5, Claude 4, Gemini 2, Mistral, or local models via Ollama. OpenClaw is model-agnostic — any model with a chat completion API works. The LLM Router skill lets you dynamically choose models based on task complexity and cost. Layer 2: The Agent (OpenClaw Core) The runtime that manages the conversation loop, tool execution, memory, and planning. This is OpenClaw's ~430,000-line codebase. It receives user input, reasons about which tools to use, executes them, and synthesizes the output. Layer 3: Tools (Function Calling) Tools are functions the model can invoke during reasoning. Examples: `readFile`, `searchWeb`, `runCommand`, `callAPI`. These are low-level building blocks — individual actions the agent can take. Layer 4: Skills (Packaged Capabilities) Skills are curated bundles of tools + configuration + prompts that solve a specific use case. A skill like RAG Pipeline bundles document ingestion, embedding generation, vector search, and response synthesis into a single installable package. Skills are defined by SKILL.md manifests. Layer 5: Output The final result: generated text, modified files, API calls, browser actions, or system commands.
This distinction matters because it's the source of most confusion in the ecosystem. Tools are verbs. They're individual actions: read a file, make an HTTP request, run a shell command, generate an embedding. Tools are low-level, generic, and reusable. Skills are nouns. They're packaged solutions: 'RAG Pipeline,' 'Code Reviewer,' 'Gmail Connector.' Skills combine multiple tools with configuration, system prompts, and workflow logic to solve a specific problem. A practical example: Without skills (tools only): You tell OpenClaw to research a topic. The agent uses the `searchWeb` tool, reads results with `readPage`, writes notes with `writeFile`, and synthesizes with the model. You have to orchestrate this manually or rely on the model's reasoning. With the Deep Research skill: You tell OpenClaw to research a topic. The skill handles everything: multi-source search, content extraction, fact verification, citation tracking, and synthesis into a formatted report. The skill's SKILL.md defines the workflow, system prompt, and quality checks. The value proposition: Skills turn OpenClaw from a generic tool-calling agent into a specialized, reliable worker. That's why the skills directory has 5,700+ entries — each one encapsulates expert knowledge about how to solve a specific problem.
Every OpenClaw skill is defined by a SKILL.md file. This is the contract between the skill developer and the OpenClaw runtime. A SKILL.md contains: Metadata: Name, version, author, description, category, tags. System prompt: The instructions the model receives when the skill is active. This is where the skill's 'expertise' lives — detailed instructions on how to handle specific tasks. Tool declarations: Which tools the skill uses and how they're configured. A skill can use built-in tools (filesystem, network) or declare custom tools with typed schemas. Permissions: What the skill is allowed to access — files, network, system commands, APIs. Note: in OpenClaw, these are advisory rather than enforced. For enforced permissions, see NanoClaw. Dependencies: Other skills or packages required. Skills can depend on other skills, enabling composition. Configuration: User-configurable settings like API keys, output formats, default behaviors. To learn how to create your own skills, see our comprehensive creation guide.
The Model Context Protocol (MCP) is Anthropic's open standard for AI tool interfaces. It's rapidly becoming the universal way for AI models to interact with external tools — not just in OpenClaw, but across Claude, GPT, Gemini, and every major agent framework. MCP vs OpenClaw's native tools: OpenClaw's built-in tools are proprietary to the OpenClaw runtime. They work great within OpenClaw but don't work anywhere else. MCP tools (called 'servers') use a standardized protocol. An MCP server built for Claude Desktop works with OpenClaw (via adapter), Nanobot (native), and any other MCP-supporting agent. MCP vs Skills: MCP servers are closer to tools than skills in the OpenClaw hierarchy. They provide individual capabilities (read a database, search the web, manage files) but don't include the workflow orchestration, system prompts, and configuration that skills provide. Think of it this way: MCP servers are universal tools. Skills are OpenClaw-specific recipes that use tools (including MCP servers) to solve problems. For a detailed comparison with implementation examples, see our MCP vs Skills deep-dive.
With 5,700+ skills in the directory, quality varies. Here's our evaluation framework: Security: Does the skill request appropriate permissions? Does it access more than it needs? Check the SKILL.md permissions section. For critical workloads, follow our security checklist. Rating: Our directory shows community ratings (1-5 stars) based on reliability, documentation quality, and maintenance activity. Verification status: Verified skills have been through our security review process. Community skills have basic automated scanning. Unreviewed skills should be manually inspected before use. Version and maintenance: When was the skill last updated? Is the author responsive to issues? Skills that haven't been updated in 6+ months may have compatibility issues with newer OpenClaw versions. Compatibility: Does the skill work with your model? Some skills are optimized for specific models (e.g., GPT-5's function calling vs Claude's tool use). Check the SKILL.md for model requirements. For a guided skill selection process, see our how to choose skills guide.
Are skills just plugins by another name? Essentially, yes. Skills are OpenClaw's plugin format. The term 'skill' was chosen to emphasize that each one encapsulates expert knowledge, not just code. But functionally, skills are plugins. Can I use MCP servers without installing them as skills? Yes. OpenClaw supports mounting MCP servers directly via the `--mcp` flag. They run as external processes and communicate via the MCP protocol. This doesn't require a SKILL.md. What happens when two skills conflict? OpenClaw uses a priority system. Later-loaded skills take precedence. If two skills define the same tool, the agent uses the higher-priority version. For complex setups, the Agent Orchestrator manages skill coordination. Can a skill use other skills? Yes. Skills can declare dependencies on other skills. The RAG Pipeline skill, for example, depends on the Embeddings Manager for vector operations. Is it better to use MCP or native OpenClaw tools? For portability across agents: MCP. For deep OpenClaw integration: native tools. For most users: whichever has the better implementation for your specific use case.