dsh-mcp-proxy

By default, dsh-mcp-client registers every tool from every configured server onto ctx.tools. A handful of servers quickly means hundreds of JSON schemas in the system prompt of every single request, whether or not the model ever calls one.

This plugin replaces all of that with exactly two tools: mcp_discover searches the tool catalog by keyword and returns each hit's server, name, description, and argument names; mcp_call invokes a specific tool by server and name, passing arguments through unchanged. Regardless of how many servers or tools are configured, the standing prompt cost stays constant. Connections are established lazily, nothing connects at boot, and the catalog is cached on disk so mcp_discover answers from a cold start without reaching any server. If one server is temporarily unreachable, discovery degrades gracefully with a named reason instead of failing outright, and tool-level failures come back as a structured flag rather than prose to parse.

If a deployment wires up a dozen MCP servers but only a few tools see regular use, this proxy shape keeps the prompt lean and defers connection cost until the model actually asks. It also runs cleanly alongside dsh-mcp-client: promote the high-frequency tools to native visibility so the model calls them without a discovery round trip, and leave the long tail behind the proxy. Tool names do not collide between the two. This suits multi-server, low-frequency-call, prompt-budget-sensitive setups.