Fetch a document and write it into another system as code, and the content is a variable in memory rather than something serialized into context twice. The three discovery tools in front of it are deliberately graduated: names first, semantic search when the names are not clear, and full schemas only for the tools you actually picked.
A meta-server that connects to your MCP servers and re-exposes them through four tools: three for graduated discovery, and one that runs JavaScript calling the underlying tools directly.
- get_mcps_servers_overview — tool names only, across every connected server, with no schemas loaded
- semantic_search_tools — describe the task in plain English and get ranked matches with similarity scores
- get_tools_overview — full definitions for just the tools you name, with parameter schemas and example usage
- run_functions_code — code that calls several tools in sequence or parallel, keeps intermediate data in memory, and returns only the final result
- Tool metadata and analyzed output schemas persisted across sessions
Node.js with npm or yarn, plus the MCP servers you want to aggregate. The README's run line names @dbestai/mcp-of-mcps, which npm does not have; the server list is passed inline with --config or as a file with --config-file. Version 3.1.0.