capa

CAPA is a CAPA (Capabilities Manager) MCP server that proxies and loads dynamic tools for AI agents. In this setup, CAPA exposes a Brave web search tool as an MCP-capable service, allowing your agents (Cursor, Claude, or other MCP clients) to discover, load, and execute the Brave-based search capability through a unified capabilities.yaml-defined workflow. You can initialize projects, define skills and tools, and then run the available tool commands directly from the MCP interface or via the capa sh CLI, enabling on-demand tool usage with credential and security controls managed by CAPA.

Prerequisites:

• Node.js (for npx) installed on your machine or environment where you run the MCP server
• Internet access to fetch the @modelcontextprotocol/server-brave-search package via npm/yarn/npx

Installation steps:

1. Ensure Node.js is installed. Verify with: node -v npm -v

2. Run the CAPA Brave search MCP server via npx (as defined by the server configuration): npx -y @modelcontextprotocol/server-brave-search This will install and start the MCP server component that provides the Brave search tool.

3. Configure environment variables if needed. For example, set the Brave API key required by the tool: export BRAVE_API_KEY="<your Brave API key>"

4. If you are integrating CAPA into an MCP project, include the server in your capabilities.yaml under servers with the appropriate def.cmd and def.env as shown in the mcp_config example.

Tips:

• The Brave search capability is exposed as an MCP tool. You can invoke it via the CAPA CLI or through your MCP client by referring to the server id (capa) and the tool name (brave-web-search as defined in the capabilities.yaml example).
• If your environment requires credentials, manage them through CAPA's interactive UI or via environment variables in your deployment setup.
• Ensure network access to the Brave search service and that the API key (BRAVE_API_KEY) is kept secret and rotated as needed.
• This server uses npx to fetch and run the package; for production, consider bundling or pinning the package version to avoid breaking changes.