The Agent GTM Window Is Now
Here’s a number that should make every infrastructure builder sit up: across 2 million tracked companies, 1,412 have published MCP servers. That’s 0.07% penetration.
Publishing an MCP server is free. Listing on registries is free. Adding /.well-known/agent.json to your domain is free. And most vertical niches — the specific tool categories agents actually search for — are wide open.
This is the window. Let me show you why it’s closing.
The acceleration curve
Six months ago, roughly 56 new MCP servers were launching per month. Last month, 301. That’s a 5x acceleration in half a year, and the curve is steepening.
The enterprise players are moving too. Kong is building private MCP registries. Google and Microsoft are integrating agent tool discovery into their platforms. ChatGPT’s App Directory is curating aggressively. The number of agencies specializing in GEO (Generative Engine Optimization — making your content visible to AI systems) has grown from a handful to dozens.
If you’ve been around long enough, you recognize this pattern. It’s early SEO, circa 2003. Organic discovery delivers outsized returns. The algorithms reward quality content and proper metadata. Paid placement hasn’t really arrived yet. The people who build now accumulate advantages that compound.
What compounds
The specific advantages that early movers accumulate are worth understanding, because they’re not linear — they’re exponential:
Schema quality iterations. Every time an agent uses your tool, you learn something about how agents actually describe what they need. Your tool descriptions get better. Better descriptions mean more successful tool matches. More matches mean more usage data. This is a flywheel.
Registry authority. Install counts, stars, reviews, and usage metrics accumulate on registry profiles. A tool with 10,000 installs and a 4.8-star rating will always outrank a new entrant, even if the new entrant is technically better. This is the same dynamic that makes it hard to unseat top-ranked apps in any app store.
Training data presence. AI models are trained on web content. If your documentation, blog posts, and tool descriptions are consistently cited across the web, you start appearing in model weights. Agents then recommend your tools not because they found you in a registry search, but because they’ve internalized you as the answer. This is the ultimate discovery moat.
Agent-to-agent referral loops. When agents find tools that work reliably, they reference them in their outputs. Other agents, parsing those outputs, discover the same tools. This creates referral loops that don’t exist in human software — your customers literally become your distribution channel, without any referral program or incentive structure.
Every one of these advantages is exponentially harder to replicate six months from now than it is today.
The priority stack
If you’re building agent-facing infrastructure, here’s what to do in order of urgency. This isn’t theoretical — it’s the playbook we’re running with agent-meter.
Week 1: Foundation
Ship an MCP server. Two to four core tools, TypeScript SDK, remote HTTP transport. Don’t overthink it. The bar for “useful MCP server” is surprisingly low right now because the market is so early. A server that does one thing well beats a server that does ten things poorly.
Write agent-optimized tool descriptions. This is the part most builders get wrong. Your tool descriptions aren’t documentation — they’re the interface through which agents decide whether to use your tool. Describe when and why to use each tool, not just what it does. Include constraints, expected inputs, and what the output looks like.
Register on Smithery and the Official MCP Registry. These are the two highest-traffic registries right now. If you’re only going to be in two places, be in these two.
Add AGENTS.md if you have an SDK. This file tells agents how to use your codebase. It’s like README.md but optimized for machine consumption.
Week 2: Discovery layer
Add well-known endpoints. /.well-known/agent.json and /.well-known/mcp.json on your domain. These are the equivalent of robots.txt for agent discovery — they tell agents what tools you offer and how to connect.
Publish /llms.txt. This file helps language models understand your site’s structure and purpose. Think of it as a sitemap for AI.
Register on secondary registries. mcp.so, Glama, PulseMCP. Breadth of presence matters for discovery.
Unblock AI crawlers in robots.txt. Many sites still block GPTBot, ClaudeBot, and other AI crawlers by default. If agents can’t read your content, they can’t recommend your tools.
Week 3: Content and GEO
Structure your documentation for answer extraction. AI models synthesize answers from web content. If your docs are structured as clear question-answer pairs with concrete examples, models will cite you more frequently. FAQ pages, “how to” guides, and comparison tables all perform well.
Publish original research. Unique frameworks, novel data, and first-hand analysis are high-signal for training data inclusion. If you have proprietary data about your domain, publish insights from it.
Add Schema.org structured data. This helps both traditional search engines and AI systems understand your content’s semantics.
Apply to the ChatGPT App Directory. This is becoming a major discovery channel. The curation bar is rising, so apply now rather than later.
The numbers tell the story
0.07% penetration means 99.93% of the market hasn’t built agent interfaces yet. A 5x acceleration in server launches means competition is arriving, but we’re still in the early exponential phase — most of the growth is ahead of us, not behind.
The historical parallel to early SEO isn’t just an analogy. The mechanics are the same: a new discovery layer emerges, organic presence delivers outsized returns, early movers build compounding authority, and eventually paid placement and saturation squeeze out the late arrivals.
We know how this story ends. The question is whether you’re building during the window or after it closes.
The window is now. Ship the MCP server. Write the tool descriptions. Register on the indexes. The compounding clock is running.