@us-all/mcp-toolkit

The 4 patterns that survive in production across 6 MCP servers.

Token-efficient defaults, pluggable error redaction, declarative aggregation, and tool-discovery meta — extracted from 6 consumer servers. Single evolution point. Cascade-automated to all consumers.

When to use this

• You're starting a new MCP server and want token-efficient defaults from day 1.
• Your existing server is past ~100 tools and clients (Cursor, Claude Desktop) start struggling with the tool list.
• You have multiple MCP servers and don't want each to reinvent slim-projection / category toggles / error redaction / aggregation boilerplate.
• You want a search-tools meta-tool so the LLM discovers tools at runtime instead of preloading every schema.

If your MCP has fewer than 20 tools and one auth mode, you probably don't need this — just ship.

The 4 patterns

import {
  applyExtractFields,        // token-efficient response projection
  ToolRegistry,              // category toggles + search-tools
  createWrapToolHandler,     // pluggable error redaction + structured errors
  // and:
} from "@us-all/mcp-toolkit";
import { aggregate } from "@us-all/mcp-toolkit/aggregate";

1. applyExtractFields — token-efficient projection

Comma-separated dotted paths with * wildcards. Slim large reads to what the model needs.

import { applyExtractFields } from "@us-all/mcp-toolkit/extract-fields";

const dashboard = { id: "abc", title: "...", widgets: [{ definition: { type: "timeseries", queries: [...] } }] };
applyExtractFields(dashboard, "id,title,widgets.*.definition.type");
// → { id: "abc", title: "...", widgets: [{ definition: { type: "timeseries" } }] }

When wired through createWrapToolHandler, tools whose schemas declare extractFields get caller-supplied projection automatically. MCP SDK validation drops undeclared fields, so each consumer must opt in per tool schema or use an explicit passthrough schema.

2. ToolRegistry<TCategory> + search-tools

Track tools by category. Pair with <PREFIX>_TOOLS / <PREFIX>_DISABLE env vars to load only relevant categories — biggest LLM context token saver.

import { ToolRegistry, parseEnvList, createSearchToolsMetaTool } from "@us-all/mcp-toolkit";

const CATEGORIES = ["metrics", "monitors", "logs", "meta"] as const;
type Category = (typeof CATEGORIES)[number];

const registry = new ToolRegistry<Category>({
  enabledCategories: parseEnvList(process.env.DD_TOOLS),
  disabledCategories: parseEnvList(process.env.DD_DISABLE),
});

let currentCategory: Category = "metrics";
function tool(name: string, desc: string, schema: any, handler: any) {
  registry.register(name, desc, currentCategory);
  if (registry.isEnabled(currentCategory)) server.tool(name, desc, schema, handler);
}

// Always-on meta-tool: lets the LLM discover other tools at runtime
const meta = createSearchToolsMetaTool(registry, CATEGORIES);
currentCategory = "meta";
tool("search-tools", "Discover tools by query", meta.schema.shape, wrapToolHandler(meta.handler));

Real-world impact (datadog, 165 tools): default load = 25K schema tokens; DD_TOOLS=metrics,monitors = 3.8K (−85%).

3. createWrapToolHandler — error redaction + structured errors

Replaces the tools/utils.ts boilerplate that 6 consumer repos all reinvented.

import { createWrapToolHandler } from "@us-all/mcp-toolkit";

const wrapToolHandler = createWrapToolHandler({
  redactionPatterns: [/DD_API_KEY/i],   // merged with built-in defaults
  errorExtractors: [
    {
      match: (e) => e instanceof WriteBlockedError,
      extract: (e) => ({ kind: "passthrough", text: (e as Error).message }),
    },
    {
      match: (e) => e instanceof DatadogApiError,
      extract: (e) => ({
        kind: "structured",
        data: { message: e.message, status: e.code, details: e.body },
      }),
    },
  ],
});

Built-in redactions: api_key, app_key, authorization, bearer …, password, secret, token. Use the bare wrapToolHandler export for zero-config.

4. aggregate(fetchers, caveats, formatReason?) — declarative aggregation

For tools that fan out 3–7 sequential API calls into one structured response. Replaces the Promise.allSettled + per-slot caveats.push(...) block that consumers all repeated.

import { aggregate } from "@us-all/mcp-toolkit/aggregate";

const caveats: string[] = [];
const result = await aggregate(
  {
    monitor: () => api.getMonitor(id),
    state: () => api.getMonitorState(id),
    events: () => api.searchEvents({ monitorId: id, window: "30m" }),
    downtimes: () => api.listDowntimes({ monitorId: id }),
  },
  caveats,
);
// result: { monitor, state, events, downtimes } — null per slot on failure
// caveats: ["state failed: timeout", ...] — surfaces partial failures

Type inferred from the input object. Rejected slots become null. formatReason lets callers customize caveat strings.

Install

pnpm add @us-all/mcp-toolkit
# peer deps
pnpm add @modelcontextprotocol/sdk zod

Node 22+, ESM, TypeScript strict. Peer SDK: ^1.27 || ^1.28 || ^1.29.

Sub-exports for tree-shaking:

• @us-all/mcp-toolkit — main barrel
• @us-all/mcp-toolkit/extract-fields
• @us-all/mcp-toolkit/registry
• @us-all/mcp-toolkit/wrap-tool-handler
• @us-all/mcp-toolkit/aggregate

Cascade automation

.github/workflows/cascade-bump.yml — when toolkit publishes, opens an automated PR in each of the 6 consumer repos that:

• bumps the @us-all/mcp-toolkit dep pin to the new version
• bumps the consumer's PATCH version
• runs pnpm install + build + test (PR creation gated on success)
• inserts a "최근 변경사항" entry in CLAUDE.md

PRs are not auto-merged — humans gate each consumer release. Re-run safe (skips if already pinned).

Auth: GitHub App us-all-bot (per-matrix-job installation token, no PAT, no expiry, scoped install). Setup walkthrough in CLAUDE.md.

Consumer suite (proof of concept)

Six production servers built on this toolkit. All MIT, distributed via npm under @us-all/*, searchable on the MCP Server Registry under io.github.us-all/*, and listed on Glama.

| Server | Tools | npm/mo | Registry namespace | |---|--:|--:|---| | @us-all/datadog-mcp | 165 | 5,010 | io.github.us-all/datadog | | @us-all/openmetadata-mcp | 170 | 3,122 | io.github.us-all/openmetadata | | @us-all/google-drive-mcp | 98 | 3,243 | io.github.us-all/google-drive | | @us-all/mlflow-mcp | 82 | 2,717 | io.github.us-all/mlflow | | @us-all/android-mcp | 76 | 2,405 | io.github.us-all/android | | @us-all/unifi-mcp | 54 | 2,566 | io.github.us-all/unifi |

~990 LOC of duplicated boilerplate eliminated across the suite at v0.1.0–v1.0.0 migrations.

STANDARD.md

STANDARD.md is the why — the conventions and rationale. This package is the how. New MCP authors read STANDARD.md first; experienced ones go straight to the API.

Testing

pnpm install
pnpm build
pnpm test     # 54 unit tests

Coverage: extract-fields edges (wildcards, backtick keys, array projection), registry semantics (allowlist/denylist, search matching), wrap-tool-handler (success/error paths, custom redaction, errorExtractors), aggregate (success/reject mix, custom formatReason, concurrency).

Contributing

See CONTRIBUTING.md. New shared patterns belong here — single source of truth for the @us-all suite.

License

MIT