dsh-mcp-lazy

dsh-mcp-lazy

A lazy MCP gateway for [DeepSeek Harness](https://github.com/deepseek-ai). It puts one

tool in front of the model instead of N MCP tool schemas: servers start on first use, idle

out again, and their tool metadata is cached on disk so search and describe never spawn

anything.

Why

@deepseek-ai/dsh-mcp-client connects every configured server at startup and registers every

one of its tools as a native tool. As its own README puts it, tool descriptions and input

schemas "enter every request while the tools are registered". A handful of servers therefore

costs thousands of tokens on every request and one resident child process each — whether or

not the model ever calls them.

This plugin keeps the model-facing surface at exactly one tool whose schema never changes,

discovers tools on demand from a disk cache, and only spawns a server when a call needs it.

The numbers

Measured against chrome-devtools-mcp@1.6.0 (29 tools), both sides rendered the same way —

JSON bytes of the tool definitions, then four bytes per token:

| | per request |

| --- | --- |

| native registration | 21252 bytes ≈ 5313 tokens |

| this gateway | 1525 bytes ≈ 381 tokens |

| saved | 92.8% |

The gateway figure is constant: configuring ten more servers does not move it, because their

schemas are read on demand rather than sent every request.

The honest counterweight. The gateway costs a fixed 1525 bytes, so it wins only when a

server's rendered tool definitions exceed that. The bundled fixture offers 7 small tools and

the saving drops to 1.7%. A server with a couple of tiny tools would make the gateway a net

loss. Measure your own before assuming:

pnpm run measure:savings                                          # local fixture
node scripts/measure-token-savings.mjs --npx 

Install

dsh plugin --profile  add dsh-mcp-lazy

Works with the DeepSeek Harness 0.1.5-rc and 0.2.0-rc lines (verified against

0.2.0-rc.2); the package's peer ranges declare exactly that.

That installs the package into the profile and registers its bundle patch, which inserts the

mcp-lazy row. Then give it your servers in the profile's cordis.patch.yml:

- id: mcp-lazy
  config:
    servers:
      - serverName: chrome
        transport: stdio
        command: npx
        args: ['-y', 'chrome-devtools-mcp@1.6.0']
        lifecycle: lazy
      - serverName: docs
        transport: streamable-http
        url: http://127.0.0.1:3000/mcp

Restart the profile. Every field is documented in [docs/configuration.md](https://github.com/wings1848/dsh-mcp-lazy/blob/main/docs/configuration.md).

Moving servers off @deepseek-ai/dsh-mcp-client

Every writer of MCP configuration in this ecosystem emits a @deepseek-ai/dsh-mcp-client row:

the config-manager panel hardcodes that package name, @hyzyn/dsh-codegraph writes a managed

row, and a hand-written config follows the same convention. Such a row registers each MCP tool

as a real tool, so its schemas enter every request — and a server listed in both places

cancels the saving this plugin exists for, with no error and nothing to notice. (The mcp({})

status output does warn you — for any server the other plugin serves natively, whether or not