The Agent Ecosystem Is the New App Store: MCP Opportunities for Solo Founders

Agents Don't Have a Browser

Every distribution channel you know was designed for a human visitor: search results to rank in, ads to click, landing pages to convert, forms to fill. AI agents have none of that apparatus. An agent doesn't "visit" your product — it either has a tool that calls your capability, or it doesn't. If your service isn't reachable as a callable tool, it doesn't exist for the fastest-growing class of software consumer.

That is why the Model Context Protocol matters more to a solo founder than its plumbing suggests. MCP is an open protocol, introduced by Anthropic in November 2024, that standardizes how models discover and invoke external tools — the USB-C-of-AI analogy got worn out because it's accurate: one integration standard, any client. OpenAI adopted it for their agents stack in 2025, Google followed for the Gemini ecosystem, and by 2026 supporting MCP is table stakes for every serious assistant, IDE, and coding agent. When every major platform converges on one protocol for tool calls, the protocol becomes the new distribution surface — and protocol-level distribution has a property indie builders have never had before: you ship once, every assistant can use it.

Compare that to the old app-store model. Getting distribution used to mean per-platform SDKs, review queues, rev-share, and native builds per ecosystem — a tax that made one-person companies structurally uncompetitive on platforms. An MCP server is a plain JSON-over-stdio/HTTP service: one implementation, no review process, no rev-share, instantly compatible with Claude, ChatGPT, Gemini, Cursor, Claude Code, Codex, and the long tail of agents built on top of them. For the first time, "write the tool once" is literally the entire porting strategy.

What the Ecosystem Is Missing

The ecosystem's shape is visible in what's trending. In our September report research, 8 of the top 15 GitHub trending repositories were tools for the AI-agent development workflow itself — not agents, but picks and shovels: agent version control, harness optimization, output coordination. The gold rush metaphor holds: the reliable money so far is infrastructure for the people running agents, and that's where the open openings cluster.

Here's the current map, with the gaps a solo founder can actually fill:

Read the table with one question: which of these gets more valuable as the big platforms standardize further? Plumbing commoditizes (that's what protocols do); judgment, trust, and cross-vendor aggregation do not. The defensible position is never "I connected the API" — it's "I curated the registry," "I maintain the test suite," "I encoded the domain rules."

The Pattern That's Already Working

The proof that protocol-level distribution pays isn't hypothetical — it's visible in what solo builders already shipped through it, and in the demand signals around September's report:

• Agent workflow tools are the hottest category on GitHub. The trend list is dominated by picks-and-shovels:

  ponytail

  (1,354 stars) coordinates agent output,

  atlas

  (888 stars) does version control for agents, and harness-tuning tools pulled 500+ stars each. When infrastructure repos out-trend the agents they serve, the market is telling you where the budget is.
• Spend-and-budget guards are an explicit gap. One of the eight opportunities in our September report was a cross-provider token budget guard — per-project budgets, threshold alerts, cost attribution per agent — driven by founders discovering runaway loops on their invoice. A tool that talks MCP and tracks what every tool call costs is infrastructure every multi-agent team will need before they're comfortable at scale.
• The $149 e-commerce check. An Indie Hackers founder launched a $149 audit showing merchants whether their storefront "breaks" AI shopping assistants. It sold because it converts an invisible risk (agents can't parse your store) into a visible verdict. The same wedge exists at the protocol layer: "does your tool surface work correctly for agents" is the 2026 version of "is your site mobile-friendly," and barely anyone offers the check.
• Open-sourcing the tool, selling the insight. The most repeatable solo pattern in this ecosystem: ship a focused MCP server as open source, let protocol-level distribution carry it into every assistant for free, and monetize the layer above it — hosting, dashboards, compliance, or the curated data the tool needs. Open source is the distribution strategy here, not the product.

The Platform Risk You Must Price In

Not everything in this ecosystem is a durable business, and the cautionary tale is recent enough to still have lesson-energy: a solo founder grew a Codex companion app to $9K/mo — and then OpenAI shipped the feature natively. A year of traction, erased by one release from the platform whose surface you built on. Our September report flagged platform risk as the defining risk of wrapper businesses, and it applies to MCP tooling with one twist.

The twist: MCP commoditizes the connection layer, which means anything that is merely a connection is now a commodity by construction. If your product's entire value is "this API, but reachable by agents," then the API owner can ship your product as a config file. What survives platform risk is everything that isn't the connection:

• Aggregation across providers — a budget guard covering Anthropic, OpenAI, and Google is harder to Sherlocks than one covering a single vendor.
• Curated data and registries — the value is the editorial judgment, not the pipe.
• Domain expertise — a compliance or bookkeeping server whose value is knowing the rules survives the protocol getting easier.
• Trust infrastructure — security review, uptime history, audit logs; nobody trusts "the official one" blindly, they trust the independent layer.

A practical test before you build: the config-file test. Ask whether the platform vendor could reproduce your product with a config file and a weekend. If yes, you're building on rented land — fine for learning, fatal for a company. If your answer includes curation, cross-vendor scope, domain rules, or trust data, you own something.

What to Do This Month

1. Ship one small MCP server for a workflow you personally run. Not a business — a weekend exercise. The fastest way to understand protocol-level distribution is to watch an agent you use daily start calling your tool. Your second server will be better informed than a hundred hours of research.
2. Audit your own agent stack for the missing tool. Write down every time this month an agent did something you wished it couldn't, or failed to do something you wished it could. That list is unmet demand collected from your own workflow — the most honest market research available.
3. Cross the trust gap deliberately. When you publish, publish as if a stranger must decide in 60 seconds whether to hand your server their credentials: a README with permissions explained, changelog, versioned releases, an example audit log. Trust infrastructure is individually cheap and competitively decisive.
4. Price in platform risk from day one. Run the config-file test on your idea before writing code. If you flunk it, reposition up the stack — from connection to aggregation, curation, or compliance — until you pass.

The Bottom Line

Every era of computing has produced exactly one distribution surface that small builders could actually reach: first shareware floppies, then the web, then the app stores — each more closed than the last, until the app-store tax made one-person companies structurally uncompetitive on mobile. The agent ecosystem is the surface that cycle just produced, and for once it's open by design: one protocol, every assistant, no gatekeeper, no rev-share.

The openings are unglamorous — registries, spend guards, test suites, trust layers — but that's what picks-and-shovels always look like. The window matters too: registries and trust infrastructure ossify early, and whoever becomes the named source for a vertical's tools becomes hard to displace. The founders who treat MCP as "just plumbing" will watch the ones who treated it as the new app store compound distribution they can't buy later.