A growing share of software choices now happen inside an agent. When someone asks Claude, ChatGPT or a coding agent to "find the CTOs at these 40 companies" or "transcribe this call", the agent searches the tools it can reach, reads their descriptions and calls one. The person often never sees your homepage.

The play: list your API or product as a per-call tool wherever agents find and pay for tools, treat the first paid call as activation, then offer a subscription to the accounts that keep coming back. The honest question is whether it can really increase distribution, or whether it just adds a new way to bill people who would have found you anyway.

The friction it removes is real. An agent user doesn't want to sign up, add a card and paste an API key for every service a task touches. Composio Instant puts paid tools behind one prepaid wallet: calls are "paid per use from your Composio wallet" and the FAQ says "Instant provider usage has 0% markup". Its sample prices, which the page labels illustrative, are about $0.005 a search, $0.020 an email lookup, $0.090 a voice call and $0.003 a scrape. Composio launched Instant on 8 October with "over 100 paid tools" and a $2 starting balance per account. Pro is $29 a month as of 25 October 2026.

The same shift is happening at protocol level. Apify made over 20,000 Actors payable via x402 in June, so an agent can pay for a run with no Apify account. Stripe now documents how to accept x402 payments from agents and record them as ordinary PaymentIntents. Its example charges $0.01 per request.

The shelves are crowded. An independent count of the official MCP Registry API found 30,375 unique servers as of 10 September, about three times the May figure, with 16,356 publishers listing a single server. Smithery shows 27,365+ MCPs. Apify Store lists 97,321 Actors and Zapier MCP covers 9,000+ apps.

Demand exists, but it is concentrated. x402.org reports, for the last 30 days, 75.41M transactions worth $24.24M from 94.06K buyers to 22K sellers. Our arithmetic: that averages about $0.32 a transaction and just over four buyers per seller. Earlier, Visa's Cuy Sheffield put adjusted x402 volume at ~$19M, and a joint Visa and Artemis analysis found about 4,000 wallets making up roughly 90% of adjusted spending. Apify says it paid out $1.7M last month. Spread across 97,321 Actors that is about $17 each, and since a mean like that hides a long tail, most listings earn far less. On Smithery the top listing shows 75.56k uses while most of the front page shows a few thousand or fewer.

Selection is decided by text, and it is fragile. Claude's tool search looks at tool names, descriptions, argument names and argument descriptions, returns five matches by default, and Anthropic notes that picking the right tool "degrades once you exceed 30–50 available tools". OpenAI says plainly that ChatGPT decides when to call your tool "based on the metadata you provide". A University of Maryland study found that edited descriptions got over 10 times more usage from GPT-4.1 and Qwen2.5-7B, and that between two identical tools GPT-4.1 picked the first one 80.2% of the time. The authors suggest adding signals based on a tool's real track record. Anthropic's directory already ranks by usage, and every entry is eligible for Suggested Connectors, the in-chat recommendations shown when a task fits.

New users or a new billing path? Both, depending on who arrives. Composio says "your own accounts come first", and its page says connected accounts aren't charged by Instant. So your existing customers keep paying you directly, and wallet calls come from people who had no account with you. That is new reach. Neither page says providers learn who those callers are, though, so the relationship sits with the marketplace until you give the user a reason to come to you. The best-documented indie result we found, a developer with 98 Actors and 2,500 users, credits Google ranking his Store pages and cross-discovery between his own listings. He doesn't mention agents.

Verdict. Listing increases distribution when your tool is the best answer to a common agent task that has few strong alternatives, when your description states that task in the user's own words, and when the platform ranks or suggests by real usage, so early calls compound. It does little for commodity APIs (generic search, weather, scraping) in crowded categories, for tools that need setup an agent can't do, or where the platform keeps the customer and you only see a payout line. For most solo builders it is a cheap extra channel worth testing for 60 days, and a poor main acquisition plan.

Keep it factual. Anthropic's directory policy requires descriptions to "precisely match actual functionality" and bars coercing Claude into calling software. The study's "most effective function" line is the kind of claim that policy rules out.

Illustrative model, made-up numbers. 5,000 agent tasks a month in your category on one platform. Your tool reaches the top five results in 30% (1,500). It is chosen in 20% of those (300 sessions) from 120 distinct accounts. 25 accounts come back on three or more days, and 3 subscribe at $19. Per-call revenue at $0.01 is about $3. Subscriptions bring $57. Per-call revenue barely registers; the repeat accounts carry the value, so the steps that turn a caller into a known user matter most.

From today's Daily Briefing →