The monetization layer for the agent economy

Every wave of software found its business model late. The web ran on hope until search ads. Mobile ran on paid downloads until in-app mediation. Agents are at that same moment right now: thousands of MCP servers and agent tools doing real work for real users — weather, flights, code review, PDFs — and almost none of them earning a cent.

We built Lulu Ads to be that missing layer.

The mechanism

A publisher adds one line to their server:

mcp.add_middleware(LuluAdsMiddleware())

From then on, tool responses can carry one extra, clearly labeled data field — sponsored — matched to the stated intent of the request. Not injected into the prompt. Not an instruction to the model. Just data, disclosed as what it is.

The host model — Claude, GPT, Gemini, whatever runs the conversation — decides on its own judgment whether that offer is relevant enough to show. If it renders and converts, the publisher earns 70%. If it doesn't render, nobody pays and nothing breaks.

Why this is the right shape

Chat interfaces have owners with sales teams. The tool layer doesn't — it's tens of thousands of independent developers with traffic and no monetization option that isn't gross. Every alternative we studied either injects instructions into prompts (a trust violation the ecosystem will reject), or sells impressions (an invitation to bot fraud), or requires owning the UI (which MCP servers don't have).

Ours requires none of that. A 150ms fail-open budget means the ad layer can never slow a tool down. A quality gate means junk never serves. CPA-only pricing means fraud attacks a price we don't charge.

The agent economy is going to need an economic engine. We intend to be it.