A-01 General arrangement
Make an APIspeak agent.
Build MCP servers around the work an assistant actually needs to do. Anthropic's MCP Builder takes your coding assistant through research, implementation, review and evaluation in one practical development guide.
One tool, exploded.
An MCP server sits between your service and an assistant. The guide treats each tool as an assembly of deliberate parts, and the quality of the server depends on how well an assistant can find, call and recover with those parts.
- Namegithub_create_issueA
Discoverable nameService prefix, action first:
{service}_{action}_{resource}, so tools stay distinct alongside other servers. - Input schemaZod (TypeScript) · Pydantic (Python)B
Typed inputConstraints, clear descriptions and examples inside the field descriptions.
- OutputoutputSchema · structuredContentC
Structured resultsStructured data alongside text, with JSON or Markdown response formats.
- AnnotationsreadOnlyHint · destructiveHint · idempotentHint · openWorldHintD
Declared behaviourHints that tell clients what a call can change.
- Context & errorslimit · has_more · next_offsetE
Focused, recoverablePaginated, relevant data and error messages that suggest the next step.
For developers turning APIs into tools.
It suits internal or external services that an assistant should be able to query and act on, built in TypeScript or Python with a coding assistant doing the work.
Built for
- Wrapping an external API or service as an MCP server, in TypeScript (MCP SDK) or Python (FastMCP).
- New TypeScript servers on SDK v2, or existing projects that already depend on SDK v1.
- Choosing between stdio for local servers and Streamable HTTP for remote ones.
- Writing a read-only evaluation set that tests whether a model can use the finished server.
Not built for
- A ready-made integration for any particular service.
- Hosted server infrastructure, deployment or API credentials.
- A guarantee that generated code is production ready.
- Use without an assistant that can write files and run your toolchain.
Bring the service. Define the outcome.
| No. | Part | Purpose |
|---|---|---|
| Supplied by you | ||
| a | API documentation and target workflows | What the tools must let an assistant accomplish. |
| b | An authorized authentication method | How the server is permitted to call your service. |
| c | A Python or TypeScript project | Including any existing SDK version, deployment and transport constraints. |
| In the package 11 original files | ||
| 1 | SKILL.md | The four-phase development guide. |
| 2 | reference/mcp_best_practices.md | Naming, response formats, pagination, transport, security, annotations, errors. |
| 3 | reference/node_mcp_server_v2.md | TypeScript SDK v2 guide with a self-check script for the built server. |
| 4 | reference/node_mcp_server.md | The equivalent guide for existing SDK v1 projects. |
| 5 | reference/python_mcp_server.md | Python / FastMCP patterns and a quality checklist. |
| 6 | reference/evaluation.md | Question design, answer verification and runner usage. |
| 7 | scripts/evaluation.py · connections.py | Python evaluation client for stdio, SSE and HTTP servers. |
| 8 | scripts/requirements.txt | Evaluation client dependencies. |
| 9 | scripts/example_evaluation.xml | Five-question format example. |
| 10 | LICENSE.txt | Apache License 2.0. |
| Packaging additions | ||
| + | PROVENANCE.md · UPSTREAM-THIRD-PARTY-NOTICES.md | Source, pinned commit, attribution and upstream-wide notices. |
| + | RUNTIME-NOTES.md · evaluation-constraints.txt | The tested Python dependency constraint for the evaluator. |
| Produced by your assistant not shipped | ||
| → | Server implementation | API client, error helpers, response formatting and pagination. |
| → | Tool contracts | Typed schemas, useful descriptions and annotations. |
| → | Build and inspection checks | Compilation or syntax checks and MCP Inspector testing. |
| → | Evaluation XML | Ten verified, read-only questions, as the guide specifies. |
Dependencies & boundaries.
No Skill-specific API key, but not free of model use.
Keep secrets out of prompts and source archives. Evaluation questions should use stable, read-only tasks. Review server behaviour and access scopes before you connect anything to real systems.
What has been checked.
Python syntax, evaluator help, example XML parsing and response extraction passed. Full server generation and a model-backed evaluation have not been tested.
The full inspection record and run evidence are on sheet A-03, Benchmark.