neoxider-mcp-hub

Running multiple MCP servers means every tool schema from every configured server stays resident in the prompt on every single turn, even when the current task touches no external tool at all. neoxider-mcp-hub collapses the host-facing surface into one flat capability_hub entry. Child servers are started over stdio only when the agent discovers a capability by intent, reviews its permissions, and explicitly enables it, then shut down again. Measured resident context drops by 93.2% without sacrificing tool-selection accuracy.

Core capabilities include: searching a lightweight catalog, inspecting a single capability for permissions and configuration, starting or stopping one trusted server, proxying a child tool call through a structured arguments object, and loading one approved local skill at a time. All model-facing JSON uses compact serialization, and the capability list is inlined in the tool description so a separate search call is unnecessary. Benchmarks on Qwen3.8-27B confirm 96.4% tool-selection accuracy at the same level as 47 static schemas, with average prompt tokens reduced to roughly a quarter of the classic setup.

This plugin is suited for anyone running several MCP servers in DeepSeek Harness or another MCP client, where most tasks touch only a small subset of the available tools. If you need permission gating, human approval for third-party additions, or a context cost that stays flat as the catalog grows, this is the trade designed for you.