
Anthropic's mcp-builder skill (60,300 installs) enforces a structured 4-phase workflow — Research → Implement → Review → Evaluate — that cut one engineer's token costs by 98.7% ($98.39 → $0.13 per 1,000 checks) across 15 enterprise MCP servers. Full install guide, Phase 4 eval framework breakdown, 11 known bugs with a practical mitigation, and a clear "when NOT to use" section.
| Phase | What happens |
|---|---|
| 1. Research | Study MCP design principles (API coverage vs. workflow tools, naming conventions, context management); read SDK docs; map out authentication, endpoints, data models |
| 2. Implement | Scaffold project (TypeScript recommended); build API client, error handling, pagination; add per-tool behavior hints (readOnlyHint, destructiveHint, idempotentHint — MCP SDK annotations that tell the client whether a tool is safe to call without confirmation) |
| 3. Review | Code quality pass (DRY, type coverage, description clarity); test with MCP Inspector (Anthropic's official local GUI for testing MCP servers interactively) |
| 4. Evaluate | Generate 10 independent, read-only, complex, verifiable XML eval questions — confirm the LLM can actually use the server before deploying |
npx skills add https://github.com/anthropics/skills --skill mcp-builder/plugin marketplace add anthropics/skills followed by selecting mcp-builder.2 The workflow assumes Node.js (for TypeScript) or Python 3.10+ is already in your environment.

ctx.elicit() signature is wrong — the reference uses invented prompt= and input_type= parameters that don't exist in the actual SDKexpress and StreamableHTTPServerTransport imports — paste-and-compile fails immediatelymax_items is deprecated, should be max_lengthtsc --noEmit after Phase 2 and let the type checker catch import gaps before Phase 3.model or allowed-tools fields), but if you're already close to the context ceiling, audit your installed skill count first.11skills/mcp-builder/skills.sh
github.com
firecrawl.dev
raw.githubusercontent.com
pub.towardsai.net
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.