Build an MCP server in 20 minutes using only AI

Executive overview

MCP servers connect AI to external tools and datasets, but most tutorials overcomplicate the concept. Dylan Davis walks through a repeatable vibe-coding workflow — reverse AI interview, blueprint generation, then prompt-by-prompt implementation in Cursor — that requires zero coding experience. The demo builds a working NASA daily-photo server that fetches images and generates GPT-4.1 summaries in under 20 minutes. The same process scales to any API or data source.

If you structure your AI interactions methodically — interview first, blueprint second, build third — you can connect AI to any tool or dataset without writing a line of code yourself.

What an MCP server actually is

• An MCP server is simply a bridge that gives AI access to external tools, datasets, and dynamic prompts.
• Tools enable actions (web search, API calls, code execution); resources provide data (CRM, email, Slack, calendar).
• Dynamic prompts extract relevant data, concatenate it, and pass a structured context to the model.
• Knowing the underlying protocol is not required to build or use one.

The demo: NASA daily space photo server

• The server calls NASA's Astronomy Picture of the Day API and returns the current image.
• GPT-4.1 generates a plain-language summary of what is visible in each photo.
• Interaction happens through MCP Inspector, a lightweight local client, rather than Claude Desktop.
• Standard I/O transport was chosen over the newer streamable HTTP because that spec was only days old and not yet stable.

Step 1 — Reverse interview with AI to create a system prompt

• Start a one-question-at-a-time interview with a fast, smart model (O4 Mini High used here) rather than typing a long brief.
• Dictation via Super Whisper speeds up input — speak your context, stop the recording, paste the result.
• The AI asks one question, your answer informs the next question, gradually steering toward a precise system prompt.
• When the AI offers examples alongside each question it makes answers easier and more accurate.
• If you cannot answer a question, feed it back to the same AI — it can suggest an answer, which you then confirm or adjust.
• The finished system prompt is placed in a ChatGPT, Claude, or Gemini project so it persists across sessions.

Step 2 — Generate a specification and blueprint

• Use a capable reasoning model (O3 here) to brainstorm MCP server ideas and identify the right API, docs, and auth requirements.
• Ask the same model to produce a full technical specification: architecture, endpoints, data flow, and dependencies.
• Paste that spec into your project (the one carrying the system prompt from step 1) to generate a numbered blueprint of standalone implementation prompts.
• Gemini 2.5 Pro produced the best blueprints in testing because its large output context window allows more detailed, self-contained prompts.
• Each prompt in the blueprint must be standalone — prompts that cross-reference each other break when fed into separate Cursor conversations.
• FastMCP 2.0 scaffolding was specified throughout because it dramatically speeds up server creation.

Step 3 — Implement prompt by prompt in Cursor

• Feed blueprint prompts into Cursor one at a time; do not paste them all at once.
• Before pasting each prompt, read it to identify which documentation or API references it mentions.
• Add the relevant docs (as text files or @-references) to the Cursor context before hitting Enter — this prevents deprecated-endpoint errors and hallucinated method signatures.
• NASA API docs, FastMCP v2 docs, MCP protocol docs, and MCP Inspector docs were all pulled in as plain-text files.
• Sonnet 4 outperformed GPT-4.1 significantly during implementation — fewer bugs and higher accuracy on first pass.
• Minor edits (inserting API keys, adjusting transport type) are the only manual steps required.

Key model choices and why they matter

• O4 Mini High for the interview: fast responses, high-quality questions without long waits.
• O3 for ideation and spec writing: strong reasoning for technical architecture decisions.
• Gemini 2.5 Pro for blueprint generation: large output window produces longer, more complete prompts.
• Claude Sonnet 4 for coding in Cursor: consistently lower error rate than GPT-4.1 in this workflow.
• Model choice at each stage is deliberate — no single model dominates every task.

Broader applications

• The same three-step framework works for any MCP server, not just NASA photo fetchers.
• The interview-to-system-prompt technique also applies to deep-research tasks outside of building tools.
• Adding relevant documentation at each implementation step is the single biggest lever for reducing errors.
• The workflow is fully accessible to non-engineers — domain knowledge about what to build matters more than coding skill.