This site is built for agents too.

My work is about a customer that can’t click: the AI agent that finds, tries, and pays for an API on its own. So this site serves that customer too. An agent can read everything here, search it, and ask my AI questions without a browser, a form, or a key.

Connect in one line

The MCP server lives at https://sparker.ai/mcp. It uses Streamable HTTP, needs no authentication, and every tool is read-only.

Add to Cursor Add to VS Code

Claude.ai and the Claude desktop app

Settings, then Connectors, then Add custom connector. Paste https://sparker.ai/mcp as the URL and leave authentication empty.

Claude Code

claude mcp add --transport http sparker https://sparker.ai/mcp

Cursor, Windsurf, VS Code, and other clients

{
  "mcpServers": {
    "sparker": { "url": "https://sparker.ai/mcp" }
  }
}

Plain HTTP

curl "https://sparker.ai/api/ask?q=What+changes+when+agents+are+your+users"

Tools

Tool	What it does
`ask_brian`	Ask a question about Brian Sparker's work, career, projects, or product thinking. Returns an answer in his voice, grounded only in his published writing and background, with source URLs. Rate limited; prefer search_site or read_page for bulk reading.
`search_site`	Keyword search over Brian's essays, frameworks, bets, newsletter, and background. Returns matching passages with titles and URLs. Free and fast.
`list_writing`	List everything published on sparker.ai, newest first: essays, frameworks, principles, newsletter issues, and bets, each with its URL, date, and summary.
`read_page`	Read one published piece as markdown. Accepts a full sparker.ai URL or a path such as /field-notes/agents-search-like-its-1999.
`get_profile`	Who Brian Sparker is: current role, background, credentials, side projects, and how to contact him.

Only ask_brian calls a model, so it is rate limited per caller. The other tools read the published writing directly and cost nothing.

Reading without MCP

/llms.txt lists every page with a one-line summary.
Add .md to any essay, framework, bet, or newsletter URL for clean markdown, for example /field-notes/agents-search-like-its-1999.md.
GET /api/ask?q=… or POST /api/ask with {"question": "…"} returns {"answer", "sources"} as JSON.
/.well-known/mcp.json describes the server for clients that discover it automatically.
It is listed in the official MCP Registry as ai.sparker/brian.

Why bother

I argue that agents search like it is 1999: terse, literal, and allergic to marketing copy. If that is true for APIs, it is true for a personal website. Serving agents here is a small version of the same design problem I work on every day, and a way to check that I believe my own advice. The reasoning is in Agents search like it’s 1999 and Attribution is who called.

The answers come from the same grounded assistant as the chat on this site. They only use my published writing and background, and they cite their sources. Questions are logged with a one-way hash of the caller’s IP address (used only for rate limits), so I can see what agents ask and write the pieces that are missing.