What does this skill do, and when should you use it?
This skill from Anthropic's official Claude Code plugin directory guides MCP server design and implementation. It begins by asking about the upstream connection, audience, action count, interaction needs, and authentication model, then recommends remote HTTP, MCPB, or local stdio deployment. It selects a tool-design pattern and recommends either the official TypeScript SDK or FastMCP 3.x before scaffolding or handing off to a specialized skill. It also covers testing with Claude, review requirements, and submission to the Anthropic Directory.
It asks five discovery questions and instructs the agent to load Claude's full connector documentation before giving guidance. It recommends one deployment model among remote streamable HTTP, MCP app, MCPB, and local stdio; chooses either one-tool-per-action or search-plus-execute; and recommends the official TypeScript SDK or Python FastMCP 3.x. Depending on the result, it can scaffold a remote HTTP or local stdio prototype inline, or hand off MCP app and MCPB work to build-mcp-app or build-mcpb. It also points to reference material for authentication, tool schemas, resources, prompts, server capabilities, deployment, testing, review, and directory submission.
- A developer wrapping a SaaS, REST, or GraphQL API who needs to select a remote MCP deployment model.
- A developer connecting Claude to local files, a desktop application, localhost services, or operating-system capabilities.
- A team wrapping an API with dozens or hundreds of endpoints and needing to keep the tool catalog manageable.
- A developer building an MCP app with forms, searchable lists, charts, visual previews, or live updates.
- A developer creating a personal MCP prototype before considering MCPB or remote deployment.
How do you install this skill?
- The default remote examples permit unauthenticated deployment; production use requires explicit authentication, authorization, audience validation, credential isolation, and data-flow review.
- The workflow requires fetching Claude documentation and relies on overseas services and deployment examples; access and verification may be unreliable from mainland-China networks.
- Placeholders such as upstreamApi, rankActions, and dispatch are not implemented, so the examples should not be treated as directly runnable or tested.
- Confirmation, rollback, and recovery for destructive operations must be implemented per server; the skill does not provide a complete universal mechanism.
- Network access
- Local filesystem
- MCP Server
This skill is part of anthropics/claude-plugins-official. The README provides the Claude Code command /plugin install {plugin-name}@claude-plugins-official, but the source does not identify the plugin-name for this skill or document a standalone installation command.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/anthropics/claude-plugins-official.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/plugins/mcp-server-dev/skills/build-mcp-server" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
After installing the plugin collection in Claude Code, use a trigger from the description, such as Build an MCP server that wraps our REST API, Create an MCP integration, or Expose these tools to Claude. The skill first asks about the connection, users, action count, interaction model, and authentication, then recommends the deployment, tool pattern, and framework before scaffolding or handing off.
What are this skill's strengths and limitations?
- Discovery-first guidance helps avoid premature deployment and architecture choices.
- Covers remote HTTP, MCPB, local stdio, MCP apps, elicitation, and large API tool design.
- Makes clear framework recommendations: the official TypeScript SDK or Python FastMCP 3.x.
- Includes testing, review, and directory-submission guidance for Claude.
- It will not begin scaffolding until the Phase 1 questions are answered.
- Its workflow depends on access to Claude documentation and referenced files, so restricted network or filesystem access may limit execution.
- The source provides no test suite, pricing information, or standalone installation instructions for this skill.
- Some host capabilities have caveats; elicitation support, for example, is still rolling out.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| MCP Server Builder this page ✓ Anthropic · Official | 54 · Use with care | ★ 38k | 11d ago | Apache-2.0 |
| MCP Interactive App Builder ✓ Anthropic · Official | 59 · Recommended | ★ 38k | 11d ago | Apache-2.0 |
| Copilot MCP Widget Builder ✓ Microsoft · Official | 49 · Use with care | ★ 2.7k | 3mo ago | MIT |
| REA: Reverse Engineer Anything | 59 · Recommended | ★ 238k | 11d ago | MIT |
| MCPB Local Server Packager ✓ Anthropic · Official | 56 · Use with care | ★ 38k | 11d ago | Apache-2.0 |
It explicitly compares remote streamable HTTP, MCPB, and local stdio deployment, and compares the official TypeScript SDK with Python FastMCP 3.x. It does not provide a full comparison with unrelated products or competing skills.
How did FollowSkills review this skill?
The material covers OAuth, audience validation, forbidden token passthrough, non-plaintext secret storage, prohibition on requesting credentials through elicitation, and confirmation for destructive actions. However, the default remote scaffold can run without authentication, uses 0.0.0.0 in the Python example, and leaves least-privilege scoping, complete data-flow disclosure, rollback, and general side-effect confirmation to the implementer, so points are deducted.
The phased workflow, deployment matrix, capability checks, fallbacks, cancellation guidance, error checklist, and version-sensitive claims are coherent. However, key examples depend on undefined placeholders such as upstreamApi, rankActions, and dispatch, and the selected skill has no dedicated test suite or statically verifiable CI coverage for its key paths, so the score remains conservative under the static cap.
The skill clearly addresses cloud APIs, local processes, desktop applications, UI, OAuth, tool-count thresholds, and multiple deployment models, with some non-fit boundaries. Trigger semantics remain broad, Chinese-language guidance is absent, and mainland-China reachability for Claude documentation, Cloudflare, and related services is not established, so points are deducted.
The documentation is progressively structured around discovery, deployment, tool design, framework choice, handoff, testing, publishing, and references. It includes a version field, a version ledger, examples, limitations, and repository-level Apache-2.0 license context. It lacks a skill-specific changelog, explicit maintainer responsibility, pinned dependency versions, and systematic troubleshooting coverage, so points are deducted.
The skill provides useful guidance for choosing an MCP shape, deployment model, tool pattern, and framework, plus TypeScript, Python, and Cloudflare starting scaffolds. The output is not a complete runnable project: it requires fetching external documentation and filling multiple business-specific placeholders, so substantial implementation and validation remain necessary; the score is therefore moderate and capped by static calibration.
The files provide reference paths, a version-sensitive claims ledger, verification commands, deployment checks, and Inspector/CLI test instructions, giving some auditability. This review has only static source material and contains no skill-specific test results, third-party execution evidence, or independent corroboration, so the score remains below the static maximum.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →