AI Agent Hub
Back to plugins
⚙️

dsh-mcp-loader

Workflow Updated 2026.09.10

Run the following command in DeepSeek Harness:

dsh plugin install MikotoMyWife/dsh-mcp-loader

Paste the following prompt into your AI chat to install this plugin:

Install in DeepSeek Harness by running dsh plugin install; the full source is at https://github.com/MikotoMyWife/dsh-mcp-loader, then register id mcp-loader with package dsh-mcp-loader in your profile to activate it.

About this plugin

Mounting a dozen or more MCP servers on DeepSeek Harness means every tool schema rides along in every request, inflating token cost and making it harder for the model to pick the right tool from a long list. dsh-mcp-loader collapses each multi-tool server into a single loader tool: the real tools enter the model context only after the loader is called in the current session, and a second call hides them again. Other sessions keep seeing just the loader, so loads never leak across sessions.

Core capabilities center on per-session visibility and fine-grained control: the loader is a toggle scoped to the calling session and its subagents; auto, lazy, and eager modes govern whether servers are probed at startup; hiddenTools masks already-registered tools per agent while disabledTools drops matched tools before registration. Additional engineering touches include description presets, parameter-description truncation, server-instruction injection, an optional resource bridge, ordered re-syncs to prevent stale overwrites, transport-level resilience with bounded disposal, and a scrubbed child environment that withholds credential-shaped keys while preserving proxy and NPM variables.

It suits teams and developers running three or more MCP servers in DSH who want to spare their token budget for content rather than schemas, and any workflow that needs per-agent tool visibility or rule-based tool suppression. The runtime dependency is limited to the official MCP client library with no extra weight.

Use Cases

  • Token budget bloats when many MCP servers expose dozens of tool schemas in every request
  • Each agent or session should only see the subset of tools it actually needs
  • Compress the visible tool list for the model without losing underlying capability
  • Multiple concurrent sessions use MCP tools and must not leak visibility into one another

Best For

  • Developers running three or more MCP servers in DeepSeek Harness
  • AI workflow teams that need per-agent or rule-based tool visibility control
  • AI engineers who track per-request token cost and tool-selection accuracy