|
Agent searches set another record while everything else
halved, the whole catalog was rescored on a new rubric,
and an African carrier is the only agent-native provider
on this week's board.
Week of August 10 – 16, 2026 · 26,575 providers ·
115,308 APIs indexed
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
One in sixteen searches now arrives through the MCP
server. Agent searches hit 2,768 — an all-time
high, the second record week running — and they did it in
a week when total search volume fell by more than half.
The MCP series now reads 35 → 42 → 310 → 1,027 → 518
→ 1,924 → 2,768. As a share of
everything: 0.1% → 0.3% → 0.8% → 2.7% → 0.4% → 2.0%
→ 6.3%.
|
|
This week vs last: total searches −53% to 44.2k,
keyless API −80% to 12.7k, unique terms −3.8% to
5,338, MCP searches +44% to 2,768. Footnote: the
volume that left was automated; unique terms held
near-flat and MCP hit a new high
|
The three red numbers are one story and the green one
is another. Requests are logged by tier: internal is the apis.io site calling itself and every crawler
walking those pages; keyless is somebody hitting the API directly from outside.
|
|
Last week
|
This week
|
|
|
Internal (the site itself)
|
31,048
|
31,494
|
+1.4%
|
|
Keyless API (outside callers)
|
63,838
|
12,718
|
−80%
|
|
MCP server
|
1,924
|
2,768
|
+44%
|
|
Unique terms
|
5,547
|
5,338
|
−3.8%
|
Last issue the keyless tier jumped 1,112% and I said flat
distinct terms meant that was one automated caller
changing doors rather than new buyers. It left. Keyless
fell 80%, the site's own tier didn't move at
all, and distinct search terms went 5,595 → 5,547 → 5,338 across
three weeks — a 4.6% drift while headline volume went 123k → 95k →
44k.
That is the number to plan against. Roughly 5,300 to 5,600 distinct things asked for per
week is the durable demand line here, and it has now survived
a crawler arriving and a crawler leaving without moving
much.
Concentration check: the top 100 terms
account for just 2,004 of 44,212 requests, and 1,267 of
those 2,004 are two machine callers — "Automation Preflight API" at 1,091 (down from 1,988) and "AVORA Agent Forge" at 176. Strip those two and the top 100 is 737 requests
spread across a very long tail. Demand here is wide and
shallow, not concentrated.
Underneath, the human vocabulary is unglamorous and
specific: edi (15), RAG (13), selling partner (10), web search (10), LexisNexis (11), 42crunch (8), json-rpc (8), time-series (8), Autonomous Vehicles (8). Integration plumbing and named vendors, not
categories.
On the Pro tier: 23 attempts on ratings, 4 on compare, 3 on gaps, 2 each on areas and changes. ratings has been the most-reached-for gated resource every
single week I've run this report.
🕳️ The Gaps (demand we couldn't answer)
The pipeline saw 393 raw unmet terms,
dropped 27 as noise, and kept 366. I then
re-probed all 100 published zero-result terms against the
live API this morning: 19 of them now return results and 81 genuinely do
not. These eight are from the 81.
|
|
The Gaps — top zero-result searches: Radar
Connection Tampering 4×, Reutersconnect and
Predictiveindex 3×, then Bedrock Converse API,
Bedrock Runtime API, ANVR G7, A10networks and
Envia.mx at 2× each
|
-
AWS Bedrock (6× across three
spellings) — somebody wanted the
Bedrock Converse and Runtime API
schemas and tried Bedrock Converse API Schema, Bedrock Runtime API Schema and the typo Bedrock ConverAPI Schema. Three attempts at the same thing is not idle
browsing. This is the sharpest gap of the week: the
model-invocation layer everybody is building on, and
no contract indexed for it.
-
Radar Connection Tampering (4×) — a
named Cloudflare Radar dataset. Not the company, not
the product, the endpoint. Someone knew
exactly what they wanted.
-
Reutersconnect (3×) — Reuters'
media licensing platform. Content licensing is the API
layer every AI company is currently negotiating in
private.
-
Predictiveindex (3×) — talent
assessment, in a week where Human Resources &
Payroll indexes 867 providers and 4,164 APIs on the
network and still misses this one.
-
ANVR G7 (2×) — second week running. The Dutch travel trade association's
booking-message standard. Somebody wants
a specification indexed, not a company, and
they came back.
-
A10networks (2×) — application
delivery and security.
-
Envia.mx (2×) — Mexican multi-carrier
shipping, alongside four spellings of nbnco from someone hunting Australia's national
broadband wholesaler by domain.
A model-invocation layer, a network-measurement dataset, a
media-licensing platform, a Dutch travel standard, and
Mexican logistics. Nothing about that list is a category —
every one of them is a specific thing a specific person
needed and did not get.
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 catalog stands at 26,575 providers and 115,308 APIs. APIs are up 242 on the week; providers are down 126 — the first Monday-over-Monday decline in the provider
count across the four weeks this report has tracked it,
with duplicate merges and retirements running ahead of
arrivals. A catalog that only ever grows is a catalog
nobody is maintaining.
The submission backlog finished processing this week, and
the two batch write-ups are the ones I'd read:
-
Six of Eight New APIs Ship an MCP Server — the final batch of eight, and six of them ship a
hosted MCP server next to the REST API. Not one is a
large company: a positional-astronomy engine, a crypto
funding-rate feed, a Georgian legal marketplace, a
WebRTC monitor. MCP has stopped being a strategy at
these companies and become a checkbox. Across all four
batches: 38 providers, 1,286 artifacts, zero failed
runs.
-
The API Economy Is Not Anglophone — the third batch, where six of eight declare a home
outside the English-speaking tech centres. The one to
sit with is Tvarka, a single REST API
for Lithuanian national eID authentication and
qualified electronic signatures — state identity
infrastructure as an ordinary HTTP API, from a company
almost nobody in this audience has heard of.
Two provider profiles publish over the coming week, and
they are a useful pair because they fail in opposite
directions:
-
Siemens — 152 APIs
listed, five machine-readable
contracts. Kin Score 40.9 (thin),
Agent Readiness 30.6 (agent-aware).
The five specs it does publish are 100% callable and
score 68.7 on contract quality; the gap between 152
and 5 is the whole profile (Industrial)
-
MTN Group — 119 APIs, 115 contracts at 89.6% callable. Kin Score 47.5 (developing), Agent
Readiness 51.1 — agent-native. See
below (Telecommunications)
One thing that moved under everyone's
feet: the catalog was rescored today on Kin Score rubric 0.11.0. Numbers in this issue are not comparable to numbers in
#04. Webex, last issue's subject,
reads 51.6 today where it read 58.9 — and
its 178 contracts are still 0.0% callable, which was the
actual point.
All on the apis.io blog.
⭐ Rated This Week: MTN Group — the only agent-native name
on the board
This week's angle: quietly excellent. Not the top
of the index — the provider that beat the giants at the
thing that matters.
MTN Group scores 47.5, a developing band, on 119
APIs. Africa's largest mobile operator: ~290 million
subscribers across roughly 16 markets, and MoMo, one of
the continent's largest mobile-money platforms.
Its composite is middling. Its Agent Readiness is 51.1, band agent-native — and that band is the interesting part.
|
Facet
|
Score
|
|
Developer Ergonomics
|
62.5
|
|
Contract Quality
|
60.9
|
|
Regulatory (Telecommunications)
|
58.3
|
|
Discoverability
|
57.4
|
|
Commercial Clarity
|
34.2
|
|
Operational Transparency
|
21.1
|
|
Governance
|
20.8
|
Agent-native is a gated band, not a threshold. In Agent Readiness 0.2 you need 46
points and you must clear a gate:
idempotency and a stable error envelope.
Retries are what separate an agent from a developer — an
agent that retries a payment without an idempotency key
charges the card twice, and an agent that can't
branch on a typed error retries blindly forever.
MTN clears it. Its error semantics are verified against the contract — a shared error schema $ref'd across responses, not a table of codes in a docs
page — and idempotency is documented on the mutating
surface.
Here is what that gate does to the scoreboard, all
rescored today:
|
|
Agent Readiness
|
Band
|
|
Salesforce
|
60.8
|
agent-ready — gated down from agent-native
|
|
MTN Group
|
51.1
|
agent-native
|
|
Mastercard
|
41.0
|
agent-ready
|
|
Microsoft Graph
|
38.7
|
agent-ready
|
|
Siemens
|
30.6
|
agent-aware
|
Salesforce scores nearly ten points higher and is capped,
explicitly, for missing idempotency. A Johannesburg
carrier ships the safety rail the highest-scoring provider
on this board doesn't.
And the access model is genuinely open. MADAPI at developers.mtn.com lists 221 API products across 15 African markets and
lets anyone download the OpenAPI for a product without an account. momodeveloper.mtn.com is self-serve with a working sandbox for the MoMo
Collection, Disbursements and Remittance APIs. Downloading
a carrier's contract without registering is rare.
Testing a carrier's payment rails in a sandbox
without a commercial agreement is rarer. The 89.6%
callability comes from TM Forum — 22 TMF Open APIs across
the catalog — which is the cheapest route to contract
quality a carrier has, and MTN took it.
The cheapest points on the table: governance at 20.8 and operational transparency at 21.1
are the two lowest facets on the card and together carry
25% of the composite. There is no published versioning
policy, no deprecation practice, and almost no operational
signalling — on a surface that moves money in markets
where mobile money is the banking system.
Rate-limit signalling scores a flat zero: MTN publishes no
rate-limit artifact and no RateLimit headers in its contracts, so an agent driving MoMo has
no way to know how close it is to a wall. A status page, a
deprecation policy, and rate-limit headers in the response
would move both facets and add Agent Readiness points at
the same time.
MTN got discovery, access and safety right and hasn't
published the governance layer. That is the reverse of the
usual enterprise failure, and by far the easier half to
fix second.
Bands, rubric 0.11.0: exemplar 66+ · strong 56–65.9 ·
developing 42–55.9 · thin 28–41.9 · emerging 13–27.9 ·
minimal 0–12.9. Agent Readiness 0.2: agent-native 46+
(gated) · agent-ready 34+ · agent-aware 6+ · human-only
0+.
🤖 Reproduce this yourself
The band-gate claim is one call and one comparison. Paste
into your agent against the apis.io MCP server:
get_provider(slug="salesforce", fields=["agent_readiness"])
Read agent_readiness.band_gated_from. It says agent-native — the catalog tells you, in the record, that it capped
the band and why. Then run the same call for mtn-group and compare dimensions.idempotency.
Want the rules rather than my summary of them?
get_rating_rubric()
That returns every band, every facet weight, and all 111
checks. Grade me with it.
Or read the demand data directly: apis.io/search-terms/, refreshed every Monday.
26,575 providers. 115,308 APIs. 2,768 agent searches —
6.3% of everything, an all-time high. 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.
|
