The Demand Report — Issue #10

Requests fell by a fifth, distinct questions rose by nearly a tenth, and more of them came from agents than in any week this report has run. For the first time, every line on the scoreboard moved the way you would want it to. Week of September 14 – 20, 2026 · 27,566 providers · 142,491 APIs Every Monday I read what people and agents actually typed into apis.io — the search box and the MCP server — as a demand signal for the API economy. Nobody else has this data, so nobody else can send you this email. 📈 The Demand Report For six consecutive weeks this report has said the same thing about its own headline number: volume up, vocabulary down, believe the small one. Last week distinct terms hit the lowest point of the run. This week it reversed. Total requests fell 297,000 → 229,355, down 23%, and keyless callers 199,207 → 159,227, down 20%. Distinct search terms rose 5,056 → 5,474, up 8%. Fewer requests, more different questions. That is what a week looks like when a crawler eases off and the people underneath it stay. This week vs last: total searches −23% to 229.4k, keyless API −20% to 159.2k, unique terms +8.3% to 5,474, MCP searches +133% to 4,754. Footnote: first week of the run where volume fell and vocabulary rose; both green bars are real And the agents set a record without a shopping list behind it MCP searches came in at 4,754 — the highest this report has recorded, up 133% on the week. The series: 35 → 42 → 310 → 1,027 → 518 → 1,924 → 2,768 → 2,508 → 4,533 → 1,591 → 2,041 → 4,754. The last time this line set a record, in #07, I could show you exactly what was behind it: twenty-nine templated searches for a hosted MCP endpoint, every one at 28–32 hits, one machine reading a vertical list. A week later it was gone and I called it a spike. This week there is nothing like that. Zero terms in the top hundred contain the word MCP. No uniform block of near-identical counts on the agent side. The record arrived on a week when the rest of the log got quieter and more varied, which is the opposite of what a burst looks like. I looked for my own fingerprints too, because last issue taught me to. The tag-quality pass ran again — 60 more tags on the 15th and 16th — but it calls the REST API on a key, not the MCP server, and the two new posting workflows read local files by design. I cannot find an internal source for these 4,754 calls. One honest caveat: until Saturday, a starter-tier key that reached the paid surface over MCP was being served the full paid dataset — a tier-ranking fallthrough, fixed on the 20th. There are 311 active starter keys. Some slice of this week's agent traffic may be starter callers who found that door open, and the log cannot separate them. I would rather tell you that than let a clean-looking record stand unqualified. What I will say plainly: this is the first MCP record in the run that I cannot attribute to a single caller, and it landed on the healthiest week of human demand I have measured. The resilience stack came back Last week's five-term cluster — outage, network status, flight tracking, emergency alerts, satellite imagery — returned: Term Sep 7–13 Sep 14–20 Critical Infrastructure Outage API 197 127 Telecommunications Network Status API 196 124 Emergency Management Alert API 195 122 Satellite Earth Observation API 195 121 Aviation ADS-B Flight API 195 119 613 requests, down from 978, same five terms, same tight band. One caller, still — but a caller that came back. That is the same pattern cohorts followed: a first week I flagged as a possible one-off, a second week that made it a standing ask. Somebody is checking, weekly, whether this catalog can assemble a situational-awareness stack. The answer last week was one thin API each and no Common Alerting Protocol at all. It has not changed, and they have not stopped asking. The gate, and the human list Bounces off the paid surface came down with everything else: 13,541 → 11,331, across 24 resources. cohorts fell 7,062 → 4,522 and is still the most reached-for gated resource by a wide margin; security 3,543; scopes 643, holding last week's doubling; ratings back up to 132. Above the noise, the human vocabulary was broader than usual, and led by one thing: Weather at 66 and Weather Forecast API at 40 — 106 searches for weather, the largest human term of the week by a factor of two. Then Paylocity (25), Payments (25), Shopify (21), Gold Price API (16), Stripe (15), Machine Learning Model Hosting (14), Anthropic (10), Cisco (9), Bkash (8). A note on the tier line, since I withdrew it last week. pro reads 6,654 this week. That is the tag-quality pass again — 60 reviews on the Understanding key — and I am not printing the authenticated tiers as demand until the log can tell my keys from a subscriber's. The gate bounces above are unaffected. 🕳️ The Gaps (demand we couldn't answer) The pipeline saw 904 raw unmet terms, dropped 447 as noise, and kept 204, down from 234. The dictionary crawler is still walking — it has reached the C's through G's this week: cephalopod, chiffonier, coonskin, fishmonger, formicarium, fossick. The catalog crawler is still reading operation names back at us: mergeintouser, removeroles, returntoauthor, roleassignees, auditactor. I re-probed all 100 published zero-result terms this morning. 98 still return nothing. The two that now resolve — a2agent and, improbably, codpiece — each match a single provider. Two candidates were dropped before the chart was drawn because the provider is already indexed under a spaced name: f5nginx (F5 and NGINX are both in) and the provider half of extract alarms from ciena (Ciena is in). Fourth week of that pattern. Eight rows, each checked against the live catalog by hand: The Gaps — top zero-result searches: workamajig, swingsync and yellowjacket at 4×, prosperstack and norsk helsnett at 3×, al jazzera, bmkg and croxyproxy at 2× prosperstack (3×, plus 4 as prosperstack as — 7 in all, the most-searched genuine gap of the week) — ProsperStack, retention software for subscription businesses: cancellation flows, offers, the measurement behind them. workamajig (4×) — all-in-one agency management: project management, accounting, CRM, resourcing. swingsync (4×) — connects to a golf simulator and turns session data into shot classification and benchmarks. A data product built on someone else's hardware. yellowjacket (4×) — YELLOW JACKET, US-made HVAC gauges, hoses and refrigerant recovery equipment. Third week running that somebody has looked for the machine layer: a PLC, a heating group, and now the tools that service the heating. norsk helsnett (3×) — Norsk helsenett, the state enterprise that runs Norway's national health IT infrastructure. Typed one letter short. A national public body with real machine interfaces behind it. al jazzera (2×, plus 4 across two more misspellings — 6 in all) — Al Jazeera. Six attempts, three spellings, nothing. bmkg (2×) — Badan Meteorologi, Klimatologi, dan Geofisika: Indonesia's national weather, climate, air-quality and earthquake agency, for 280 million people. Read it against the top of this issue — weather was the largest human search of the week, and the resilience stack came back — and this is the single most on-target gap in the list. croxyproxy (2×) — a free web proxy. Two national public bodies in one week — Norway's health network and Indonesia's weather agency. The non-Anglophone pattern has been a standing one in this section since #06; this is the first time both rows are governments. apifootball came back a fourth consecutive week, at 4×. No row, by policy; noted, because four weeks is a customer. If you build in one of these lanes, this is your invitation. Add your API in about two minutes: apis.io/add — or point your agent at the apis.io MCP server and let it submit for you. 🆕 New to the Index The deployed index reads 27,566 providers and 142,491 APIs. Providers are down 30 on the week; APIs are up 8,727, and as always that line counts specification files, not products — a large contract split by resource shows up as dozens. The MCP lines did something they have not done before: Last Monday Today MCP entries (all) 4,961 5,270 +309 — real servers 2,651 2,920 +269 — hosted endpoints 1,298 1,445 +147 — providers credited with MCP 2,566 2,769 +203 Scopes (providers) 2,871 2,940 +69 Security (providers) 24,423 24,419 −4 Hosted endpoints grew by 147 in a week. The previous two weeks were +51 and +44. Real servers grew by 269 against candidate manifests by 39 — the first week the real side of that ledger outgrew the declared side by a wide margin. Three weeks ago I wrote that growth in the MCP count was mostly growth in declarations. This week it was not. Scopes crossed 10.7% of the catalog. Twenty posts published last week — the heaviest run yet. Among them: Seven in Ten Gated MCP Servers Will Let an Agent In Unaided The Last 25 APIs Added to apis.io Are Increasingly Built for Agents Healthcare Is the Biggest Tag on apis.io — and the Least Agent-Ready One Missing Field Keeps 53 Agent Cards Out of the Registry Bindbee vs Deel: One Contract Against Ninety-Nine Webhook Providers Score Twice the Market and Still Skip Idempotency All on the apis.io blog. ⭐ Rated This Week: Tomorrow.io vs OpenWeatherMap — two tenths of a point, one whole band This week's angle: weather was the largest human search of the week and agents set a record. Here are the two weather APIs people mean, and what separates them for each kind of caller. Ninety-six providers in the catalog carry the weather tag. Thirteen of them score above 50. The two most people would name are these, and on the Kin Score they are effectively tied: Tomorrow.io OpenWeatherMap Kin Score 65.1 strong 64.9 strong Agent Readiness 32.6 agent-ready 25.6 agent-aware Contracts / callable 10 / 100% 14 / 100% Two tenths of a point apart for a developer. A full band apart for an agent. And the facets underneath do not agree about which is better: Facet Tomorrow.io OpenWeatherMap Discoverability 64.8 83.3 Contract Quality 69.8 78.1 Contract Governance 28.8 33.3 Operational Transparency 52.6 34.2 Developer Ergonomics 73.8 65.5 Access Clarity 84.2 67.1 OpenWeatherMap writes the better contract and is easier to find. It wins contract quality by eight points and discoverability by eighteen. Fourteen specifications, all callable, none derived by us. If you are a person with a text editor, it is the better-documented API. Tomorrow.io built the door. Its official MCP server sits at api.tomorrow.io/v4/tomorrow-weather/mcp. I sent it an anonymous initialize this morning and got the right refusal: HTTP/2 401 www-authenticate: Bearer scope="mcp:read", resource_metadata="https://api.tomorrow.io/.well-known/oauth-protected-resource/v4/tomorrow-weather/mcp" Follow that URL and there is a real RFC 9728 document naming the resource, its authorization server and the one scope it grants. A .well-known path that cannot exist returns 404, so the metadata is honest rather than a catch-all. The refusal names the scope — mcp:read — which is one detail more than most servers manage. OpenWeatherMap has no official MCP server. What the catalog holds is a candidate manifest — a tool list we derived from its OpenAPI, marked as such, with a note that says exactly what it is: "No official OpenWeather-hosted MCP server is published. Community MCP servers exist." There are at least four of those, built by individual developers, each requiring you to bring your own key. That is last issue's authorship finding in one provider: the agent door to OpenWeatherMap exists, and OpenWeatherMap did not build it. The cheapest points on the table run in both directions. For OpenWeatherMap, one hosted MCP endpoint with the refusal above moves it from agent-aware to agent-ready without touching the contract it already wins on. For Tomorrow.io, the docs link the catalog holds for its MCP server — docs.tomorrow.io/reference/tomorrow-io-mcp — returns 404 today; the server is live and its front door is not. And neither publishes idempotency on anything, which keeps both out of agent-native regardless. Weather was the largest human ask of the week. Between these two, the human ask is a coin flip. The agent ask is not. Kin Score rubric 0.22.0, both scored 2026-09-20. Bands: exemplar 66.5+ · strong 54.3–66.4 · developing 39.3–54.2 · thin 26.2–39.2 · emerging 11–26.1 · minimal 0–10.9. Agent Readiness 0.2: agent-native 38.7+ (gated on idempotency AND error semantics) · agent-ready 28.6–38.6 · agent-aware 5.1–28.5 · human-only 0–5. 🤖 Reproduce this yourself The two weather providers, side by side, no key required: find_providers(providers=["tomorrow-io","openweathermap"], fields=["slug","score","agent_readiness"]) Then the door, the way I checked it: curl -s -D - -o /dev/null -X POST https://api.tomorrow.io/v4/tomorrow-weather/mcp \ -H 'Content-Type: application/json' \ -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' \ | grep -i www-authenticate And the count that moved this week: find_mcp(install_method="Hosted endpoint", limit=1) meta.total reads 1,445, up from 1,298 last Monday. Or read the demand data directly: apis.io/search-terms/, refreshed every Monday. 27,566 providers. 142,491 APIs. Fewer requests, more questions, and 4,754 of them from agents — the first record on that line I can't explain away. See a gap that's yours? → apis.io/add The Demand Report is a weekly read of apis.io's own search + discovery signal. Forward it to someone who ships an API. Unsubscribe | 606 W 57th New York, NY 10019