---

name: tools-and-mcp description: "Guides agents using CrewAI tools, crewai_tools exports, custom BaseTool and CrewStructuredTool patterns, MCP server adapters, tool publishing, integrations, optional dependencies, and credential boundaries." disable-model-invocation: true metadata: disco-role: operating license: MIT

Tools and MCP

Use this sub-skill when work involves official crewai_tools, custom CrewAI tools, BaseTool, CrewStructuredTool, MCP servers, tool publishing, enterprise/action adapters, or deciding which tool path is safe when credentials or optional packages are missing.

Route First

• For assigning tools to Agent or Task, kickoff behavior, or crew process design, use core-runtime.
• For RAG storage, knowledge sources, vector DB architecture, or embedding setup behind search tools, use memory-knowledge-and-rag.
• For LLM provider keys, model routing, base URLs, or function-calling model compatibility, use llm-and-providers.
• For root-level CrewAI scenario selection, use the parent crewai router.

Workflows

1. Identify whether the user needs a built-in tool export, a custom tool, an MCP integration, or a publishable tool package.
2. Check tool-catalog.md for installed export families, safe local defaults, and categories that usually require network credentials.
3. Use custom-tools.md for BaseTool, @tool, CrewStructuredTool, typed output, cache, max-usage, and publishing patterns.
4. Use mcp-reference.md for Agent(mcps=[...]), MCPServerStdio, MCPServerHTTP, MCPServerSSE, MCPServerAdapter, filtering, timeouts, and cleanup.
5. Use optional-dependencies.md before instantiating tools that may need extras, SDKs, browser runtimes, API keys, or enterprise tokens.
6. Use troubleshooting.md when tool calls fail from missing credentials, schema validation, path safety, MCP transport issues, or import mistakes.

Safe Defaults

• Prefer local, deterministic tools such as FileReadTool, DirectoryReadTool, or a small custom BaseTool when API keys are missing.
• Do not instantiate tools that make network, cloud, browser, database, or LLM calls unless the user explicitly provided credentials and approved the side effects.
• Keep secrets in environment variables or runtime config; never place real tokens in tool descriptions, examples, logs, or generated skill content.
• For MCP, prefer the native mcps field on Agent for normal use and reserve MCPServerAdapter for manual lifecycle control.
• Filter MCP tools to the minimum allowed set and document local process, file-system, and remote-server security assumptions.

Bundled Script

• scripts/list_tool_exports.py safely imports crewai_tools, lists exported names and inferred categories, and performs no API, LLM, network, credential, file mutation, or tool-instantiation work.

Usability Targets

• Choose a safe, keyless tool pattern when a requested search or web tool lacks its third-party API key.
• Migrate a local MCP stdio server configuration into a CrewAI agent with tool filtering, timeout, lifecycle, and security notes.