Feature request: Implement Tool Search Tool

[bhb]

Describe your issue

Implement Tool Search Tool as described by Anthopropic here.

By adding defer_loading: true to most tools, the context required by the MCP server would dramatically shrink. This would allow us to use more of our context window for solving coding issues instead of filling it with tool descriptions.

Our current workaround is to frequently disable the MCP server and then only enable it when we know we'll need it.

MCP Config

    "shortcut": {
      "command": "docker",
      "args": [
        "run",
        "--interactive",
        "--rm",
        "--platform",
        "linux/arm64",
        "-e",
        "SHORTCUT_API_TOKEN=${SHORTCUT_API_TOKEN}",
        "-e",
        "SHORTCUT_TOOLS=stories,users,workflows,epics",
        "node:20-alpine",
        "npx",
        "-y",
        "@shortcut/mcp"
      ]
    },

Environment (please complete the following information):

• OS: Mac
• IDE: Claude Code
• Node version: v20.12.2
• Node version manager NVM

Additional context

As you can see above, we do use SHORTCUT_TOOLS, which is great, but tool descriptions still take up a large part of our context window.

[vsarunas]

Just to give some background on why this is needed, I just added the Shortcut MCP and it took up 14k tokens:

  MCP tools · /mcp
  └ mcp__shortcut__users-get-current: 50 tokens
  └ mcp__shortcut__users-get-current-teams: 60 tokens
  └ mcp__shortcut__users-list: 47 tokens
  └ mcp__shortcut__stories-get-by-id: 171 tokens
  └ mcp__shortcut__stories-search: 3.0k tokens
  └ mcp__shortcut__stories-get-branch-name: 127 tokens
  └ mcp__shortcut__stories-create: 391 tokens
  └ mcp__shortcut__stories-update: 503 tokens
  └ mcp__shortcut__stories-upload-file: 152 tokens
  └ mcp__shortcut__stories-assign-current-user: 129 tokens
  └ mcp__shortcut__stories-unassign-current-user: 133 tokens
  └ mcp__shortcut__stories-create-comment: 152 tokens
  └ mcp__shortcut__stories-create-subtask: 200 tokens
  └ mcp__shortcut__stories-add-subtask: 174 tokens
  └ mcp__shortcut__stories-remove-subtask: 136 tokens
  └ mcp__shortcut__stories-add-task: 194 tokens
  └ mcp__shortcut__stories-update-task: 247 tokens
  └ mcp__shortcut__stories-add-relation: 221 tokens
  └ mcp__shortcut__stories-add-external-link: 170 tokens
  └ mcp__shortcut__stories-remove-external-link: 170 tokens
  └ mcp__shortcut__stories-set-external-links: 190 tokens
  └ mcp__shortcut__stories-get-by-external-link: 130 tokens
  └ mcp__shortcut__iterations-get-stories: 173 tokens
  └ mcp__shortcut__iterations-get-by-id: 171 tokens
  └ mcp__shortcut__iterations-search: 1.7k tokens
  └ mcp__shortcut__iterations-create: 216 tokens
  └ mcp__shortcut__iterations-get-active: 113 tokens
  └ mcp__shortcut__iterations-get-upcoming: 113 tokens
  └ mcp__shortcut__epics-get-by-id: 170 tokens
  └ mcp__shortcut__epics-search: 2.2k tokens
  └ mcp__shortcut__epics-create: 173 tokens
  └ mcp__shortcut__objectives-get-by-id: 171 tokens
  └ mcp__shortcut__objectives-search: 1.6k tokens
  └ mcp__shortcut__teams-get-by-id: 159 tokens
  └ mcp__shortcut__teams-list: 50 tokens
  └ mcp__shortcut__workflows-get-default: 118 tokens
  └ mcp__shortcut__workflows-get-by-id: 173 tokens
  └ mcp__shortcut__workflows-list: 50 tokens
  └ mcp__shortcut__documents-create: 225 tokens
  └ mcp__shortcut__documents-list: 51 tokens
  └ mcp__shortcut__documents-search: 223 tokens
  └ mcp__shortcut__documents-get-by-id: 111 tokens
  └ mcp__shortcut__labels-list: 109 tokens
  └ mcp__shortcut__labels-create: 205 tokens

[andreasmcdermott]

@bhb @vsarunas Note that you can disable any tools you don't need as well. That will ensure they are never loaded into the context window.

See the README for instructions: https://github.com/useshortcut/mcp-server-shortcut?tab=readme-ov-file#limit-tools

---

But we will look into implementing some optimizations to the current tool descriptions, as well as tool search!

[andreasmcdermott]

@bhb @vsarunas We still haven't added the ability to load tools lazily (all though it is planned). In the mean time, we just released a new version that optimizes the tool descriptions to reduce token usage. Let us know if that helps at all (and whether you have any issues with the LLMs using the tools accurately)!