Four hours

I shipped an MCP server, a REST API, API docs, API key management, programmatic SEO pages, and a bunch of homepage tweaks for Board Paper Scraper. That was a Tuesday, after work.

Board Paper Scraper indexes NHS trust board papers. Sales teams in health tech use it to find procurement signals. The data is good, but it was locked behind a dashboard. You had to log in, click around, read things with your eyes. Fine for humans, but AI agents want endpoints.

So I built them. A REST API with Bearer token auth:

GET /api/v1/trusts/search?q=london&type=acute
GET /api/v1/trusts/:id
GET /api/v1/trusts/:id/board-papers
GET /api/v1/board-papers/:id
GET /api/v1/board-papers/:id/files/:fileId/download

And an MCP server so Claude Code, Cursor, or Windsurf can connect with a single config block:

{
  "mcpServers": {
    "board-paper-sales": {
      "type": "streamable-http",
      "url": "https://boardpaperscraper.com/api/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_API_KEY"
      }
    }
  }
}

Add that to your ~/.claude/mcp.json and you can ask "which acute trusts in London published board papers this month?" and get real answers from real data. No dashboard, no context switching.

Then API key management so customers can generate and revoke keys. Documentation with setup guides for every major AI tool. An llms.txt route so the docs are machine-readable too. Then 30+ programmatic SEO pages targeting integration and use case keywords. Then homepage improvements based on analytics.

Four hours. One PR.

The foundations matter

This was only possible because the existing codebase was well-structured. The internal API endpoints were already clean, the data layer was sensible, the patterns were consistent. AI didn't write this from nothing. It worked fast because it had good code to follow.

AI-assisted development amplifies what's already there: it extends clear patterns at speed and reproduces a mess just as quickly. The quality of the foundations determines the quality of the output.

Count the roles

The work covered backend engineering for the API, protocol integration for the MCP server, content and growth for the SEO pages, technical writing for the documentation, and product decisions driven by analytics for the homepage changes.

In a traditional setup, that's a backend engineer, a frontend engineer, a content person, a technical writer, and a product manager making the prioritisation calls. Five different roles, or one person context-switching across all of them over several weeks.

Instead it was me, after work, with Claude Code open.

Any idea can be tested

Ideas that previously didn't justify the time are now cheap enough to test.

Turning my dashboard into an API was always a good idea. It just never won a prioritisation fight. Dropping the cost from a week to four hours changed the calculation completely.

When a test takes an evening, "we haven't got time for that" starts to sound a bit thin. Try it before consigning it to next quarter. The experiment is often cheaper than another month of debate.

A caveat

I should be fair: this works best when you're building something new, or extending a clean codebase. Legacy systems with poor documentation and tangled dependencies are a different story. AI can still help, but it's slower, more cautious work.

Even then, a feedback loop that still takes weeks deserves scrutiny, and the tools probably aren't the bottleneck.