Skip to content

borghei/Claude-Skills / engineering/mcp-server-builder

mcp-server-builder

Verified

Builds MCP servers: OpenAPI-to-tool conversion, tool and schema linting, TypeScript or Python scaffolds, transports. Use when exposing an API to AI agents, building a tool server, or reviewing MCP tool definitions.

MCP Server Builder

The agent designs and ships MCP (Model Context Protocol) servers whose tools an agent picks correctly and calls safely: verb_noun names, descriptions that say when to use the tool and what it returns, typed inputs, paginated lists and gated destructive actions. Schema quality is the main lever: the model chooses tools from names and descriptions alone.

When to Use

  • Exposing an internal REST API to Claude, Cursor or another MCP client.
  • Converting an OpenAPI spec into MCP tools, then fixing what the conversion gets wrong.
  • Replacing brittle browser automation with typed tool interfaces.
  • Reviewing an existing server's tool definitions before release.

Clarify First

Before building the server, settle the inputs below. When one is missing or vague and would change the result, ask about it; when it wouldn't, use a sensible default and say which one you used.

  • Source contract: an OpenAPI spec, an existing API, or tools designed from scratch.
  • Language: TypeScript (@modelcontextprotocol/sdk) or Python (mcp).
  • Transport: stdio for local tools, Streamable HTTP for remote or shared servers (HTTP+SSE only for legacy clients).

Ask about the two or three inputs that matter most for this request, not the whole list. If the user wants a result straight away, produce it and state your assumptions at the top.

Workflow

  1. Get tool definitions. From a spec: openapi_converter.py spec.json -o tools.json (add --filter GET,POST, --exclude '/internal/*', --prefix). From scratch: write {"tools": [...]} by hand using references/tool-schema-design.md.
  2. Lint naming, descriptions and schemas with tool_linter.py tools.json. Fix every error; rewrite descriptions to say when to use the tool and what it returns.
  3. Lint interaction design with tool_schema_linter.py tools.json: add pagination to list tools, a confirm or dry_run gate to destructive ones, arrays where calls are naturally repeated.
  4. Scaffold the server with server_scaffolder.py tools.json --lang typescript|python --output <dir>, then implement each handler from references/server-implementation.md.
  5. Test with MCP Inspector and the checks in references/testing-and-troubleshooting.md.
  6. Done when both linters exit 0 with --strict, every handler returns structured errors instead of throwing, and the Inspector lists and calls every tool.

Quick Start

# from engineering/mcp-server-builder/; write output outside the repo
OUT=$(mktemp -d)
python3 scripts/openapi_converter.py assets/sample_openapi.json --prefix tasks -o "$OUT/tools.json"
python3 scripts/tool_linter.py "$OUT/tools.json"                # exit 0: no errors, 2 usage-guidance warnings
python3 scripts/tool_schema_linter.py "$OUT/tools.json"         # exit 1: delete has no confirm gate
python3 scripts/server_scaffolder.py "$OUT/tools.json" --lang python --output "$OUT/server" --dry-run
python3 scripts/tool_linter.py assets/flawed_tools.json         # exit 1: camelCase name, 1-char description, untyped property
python3 scripts/tool_schema_linter.py assets/flawed_tools.json  # exit 1: unpaginated list, ungated delete

Tools Overview

ToolPurposeKey flagsExit codes
scripts/openapi_converter.pyOpenAPI 3.x / Swagger 2.0 JSON to MCP tool definitions ({"tools": [...]}), operationId to snake_casespec, -o, --prefix, --filter GET,POST, --exclude GLOB, --no-source, --json0 converted; 2 missing or invalid spec
scripts/tool_linter.pyNaming (snake_case, verb_noun), description length and guidance, property types and descriptionsfile, --path, --strict, --json0 no errors (or no warnings with --strict); 1 issues; 2 unreadable input
scripts/tool_schema_linter.pyInteraction smells: list tools without pagination, destructive tools without a gate, chatty single-item tools, unconstrained inputsfile, --path, --strict, --json0 no warnings; 1 smells found; 2 unreadable input
scripts/server_scaffolder.pyTypeScript or Python server skeleton with one handler stub per toolconfig, --lang, --output, --dry-run, --json0 generated; 2 invalid config

The linters accept a bare tool, a tool array, or a {"tools": [...]} wrapper (converter output); use --path for other nesting. The scaffolder takes the converter output directly and derives the server name from the spec title when name is absent. Run both linters: they check different things.

Anti-Patterns

Mistake: Shipping the converter output as is. Why it matters: OpenAPI summaries describe endpoints, not when an agent should call them, so the model picks the wrong tool or none. Instead: Rewrite each description with a "Use when ..." clause and what it returns, then re-lint.

Mistake: A list tool that returns everything. Why it matters: Unbounded results flood the context window and get truncated mid-record. Instead: Cap the page size, return a cursor, and document both in the description.

Mistake: Destructive tools that act on the first call. Why it matters: The agent calls tools speculatively; an ungated delete turns a misunderstanding into data loss. Instead: Add confirm or dry_run defaulting to the safe path, and say in the description that the action cannot be undone.

Common Patterns

  • Long-running tools: emit throttled progress with a stable handle, honor cancellation, and make the final response stand alone. See references/streaming-batch-and-computer-use.md.
  • Batch-friendly tools: accept bounded arrays, take an idempotency key, return per-item results so the agent retries only failures.
  • Computer use via MCP: expose intent-level actions (navigate, click, read_page), return observable state as text first, and gate irreversible actions. For the automation engine see engineering/computer-use-automation.

Reference Documentation

Files

FilePurpose
scripts/openapi_converter.pyOpenAPI to MCP tool definitions
scripts/tool_linter.pyNaming, description and schema linter
scripts/tool_schema_linter.pyInteraction-design linter
scripts/server_scaffolder.pyServer skeleton generator
assets/sample_openapi.jsonThree-operation Tasks API (list, create, delete)
assets/flawed_tools.jsonDeliberately weak tool definitions; both linters exit 1

Scope & Limitations

Covers: tool, resource and prompt design; TypeScript and Python servers; transport choice and client configuration; OpenAPI conversion; schema and interaction linting.

Does not cover:

  • MCP clients or orchestration layers (see engineering/agent-workflow-designer).
  • Multi-agent systems that consume the tools (see engineering/agent-designer).
  • REST API design itself (see engineering/api-design-reviewer).
  • Deployment pipelines (see engineering/ci-cd-pipeline-builder).
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 468kUpdated Claude CodeCodexUnverifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 321kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 321kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 321kUpdated Claude CodeCodexVerifiedInstall