Image Search Infrastructure for AI Agents: What the Market Actually Ships

Every agent framework shipped this year can write, plan, call tools and open a pull request. Ask one for a photograph and the whole thing degrades to a shrug: a stock URL invented from memory, a broken hotlink, or an image generated because searching for a real one was harder than making one up. The models are not the weak link. The image layer underneath them is — and it is missing in a very specific, measurable way.

This is a market study, not a manifesto. Everything below was measured or read from a primary source on 18 August 2026: a census of the official MCP registry, the published terms of the four large free photo APIs, the access conditions of the two paid incumbents that shipped agent connectors this month, and a small reproducible benchmark you can re-run in a minute with no API key.

One boundary, so you can stop reading early if this is not your question: this article is about the access layer — retrieval, reachability, terms, quotas. It is not the publishing pipeline. If what you need is how to turn a draft into illustrated, credited, marked-up pages — the router prompt, the visual briefs, the assembly, the ImageObject markup — that is written up separately in the pipeline that illustrates AI-written articles.

What changes when the caller is a machine

A human searching for a photo types two or three words, scans a grid, and knows a good picture when they see one. Every stock search engine ever built assumes that loop. An agent breaks all three assumptions at once:

1. It writes sentences, not keywords. A model that has just drafted a section about mortgage paperwork does not emit mortgage; it emits “a woman signing a mortgage document at a kitchen table while a bank adviser explains the terms”. That is the natural output of a language model, and it is exactly what a token-matching index cannot serve.
2. It cannot scan a grid. Whatever the tool returns is what the agent knows. If the response is a list of URLs with no description, dimensions or credit, the agent has to fetch and look at each image — or, far more often, guess.
3. It has no hands. It cannot upload screenshots to a review queue, accept terms, rotate a key, or wait three business days for production approval. Every human step in an API's onboarding is a wall the agent stops at.

So “image search infrastructure for AI agents” is not a marketing phrase. It is a checklist: natural-language retrieval, machine-readable results rich enough to reason over, and an authentication and quota model a program can complete on its own.

The test: a sentence against a keyword index

The cheapest way to see the problem is to send an agent's own words to a keyword engine. The test below uses Openverse, the openly licensed media search run by WordPress.org, because its API is public, needs no key, and anyone reading this can re-run it in under a minute. Four briefs, of the kind a model writes when it is illustrating what it just drafted:

import json, urllib.request, urllib.parse

BRIEFS = [
  "a woman signing a mortgage document at a kitchen table while a bank"
  " adviser explains the terms, warm morning light",
  "An elderly couple enjoying a peaceful retirement together",
  "a child discovering snow for the first time in a suburban garden",
  "a delivery cyclist waiting at a red light in heavy rain",
]

def count(q):
    u = "https://api.openverse.org/v1/images/?" + urllib.parse.urlencode({"q": q})
    r = urllib.request.Request(u, headers={"User-Agent": "benchmark/1.0"})
    return json.load(urllib.request.urlopen(r))["result_count"]

for b in BRIEFS:
    print(count(b), "|", b[:48])
# 0 | a woman signing a mortgage document at a kitchen
# 0 | An elderly couple enjoying a peaceful retirement
# 0 | a child discovering snow for the first time in a
# 0 | a delivery cyclist waiting at a red light in hea

Four briefs, four empty result sets. Now cut each brief down to the two keywords a human would have typed — signing document, elderly couple, child snow, cyclist rain — and every one of them returns a full page of photographs (the API reports 240, which is where it caps the count for anonymous callers). The catalogue holds the pictures. The sentence is what it cannot parse.

Swipe the table sideways to see every column →

The Pexafy column is the same four sentences sent unmodified to search_photos on the hosted MCP server, with the latency the server itself reported. This is a comparison of query shapes, not of catalogue quality — and the catalogue is emphatically not the problem. Openverse's own statistics endpoint reports 915 million images across 52 sources on the day of the test, and it is excellent at what it was built for: keyword search, for people. Nothing about it is broken. It was simply never designed for a caller that thinks in sentences. Why tag matching breaks on a described scene — and what semantic retrieval does instead — is argued from the reader's side in searching with a sentence instead of keywords; what is new here is the measurement, and what it means when the caller is a program rather than a person.

“An elderly couple enjoying a peaceful retirement together” · 16 results in 135 ms · three of them

Rank 1Pexels Pixabay Unsplash

The market, as of 18 August 2026

Retrieval is one half. The other half is reachability: can an agent get to the catalogue at all, and on what terms? Since the Model Context Protocol became the way assistants attach to outside systems — its current specification, dated 28 July 2026, moved the protocol to a stateless request/response core so servers sit behind ordinary load balancers1 — the honest way to survey this market is to look at what is actually published, not what is announced.

The paid incumbents arrived this month

Getty Images launched its MCP server on 12 August 2026, six days before this article, exposing creative, editorial and archival search and download to AI workflows.2 It is aimed at enterprise teams and platform builders, and its own access page is unambiguous about the entry price: you need an active Getty Images agreement, such as Premium Access, plus acceptance of the Web Service terms. There is no free tier. Shutterstock content is reachable through an MCP server as well — 24 tools covering search, collections and licensing — with downloads gated behind an account that carries an active subscription.3

Both moves are rational and both are for a specific buyer. Getty booked $981.3 million of revenue in 20254 in a stock photography market sized at roughly $5.4 billion for 20265; connectors that require a contract protect that revenue rather than opening it. If you are a bank or a broadcaster, this is exactly what you want. If you are a developer wiring a side project — or an agent that must not stop and ask a human to sign something — it is a closed door.

The free libraries have not shipped one at all

Unsplash, Pexels and Pixabay are the three catalogues that most builders reach for, and none of them publishes an official MCP server. What exists instead is a shelf of community wrappers. We crawled the official MCP registry in full — 22,607 servers, 75,487 published versions — and kept every entry whose name or description is about photo or stock-image search. The pattern is uniform:

Swipe the table sideways to see every column →

import json, urllib.request, urllib.parse

def crawl(term):                      # the registry paginates with `nextCursor`
    out, cursor = {}, None
    while True:
        p = {"limit": "100", "search": term} | ({"cursor": cursor} if cursor else {})
        d = json.load(urllib.request.urlopen(
            "https://registry.modelcontextprotocol.io/v0/servers?" + urllib.parse.urlencode(p)))
        for e in d["servers"]: out[e["server"]["name"]] = e["server"]
        cursor = (d.get("metadata") or {}).get("nextCursor")
        if not cursor or not d["servers"]: return out

servers = {}
for t in ["photo", "image", "stock", "picture", "unsplash",
          "pexels", "pixabay", "getty", "shutterstock", "openverse"]:
    servers |= crawl(t)

# 137 unique servers · remote vs local is `server["remotes"]`
# the required third-party keys are in packages[].environmentVariables

Every one of those wrappers is a thin client of the same three keyword APIs, which means the benchmark above applies to all of them: the transport changed, the retrieval did not. A wrapper cannot make a tag index understand a sentence.

The clause that stops a machine is not the quota

The free libraries' quotas and terms are compared side by side, in a common unit, in the free stock photo API comparison — the approval queues, the mandatory download call, the 24-hour caching obligation, the attribution rules. That article reads them as a developer choosing a provider. Read them again as the program that has to obey them and one line stands out that has nothing to do with volume:

The same reading applies to onboarding. A demo tier that becomes production only after a human reviews screenshots of your app is not a rate limit an agent can back off from; it is a step it cannot perform at all. Pixabay states the intent explicitly — the API is for legitimate human usage, and mass automated downloading is prohibited. None of this is unreasonable: photographers give the work away and someone pays for the bandwidth. It simply describes an API whose caller was assumed to have a screen and a pair of hands.

Four requirements, and who meets them

Put the two halves together and the specification for an image layer an agent can actually use comes out short and testable. It is deliberately not a provider comparison — catalogue size, pricing and licence terms are lined up in the comparison article. These four are only about whether a program can reach the pictures at all:

1. Semantic retrieval — a full sentence returns a ranked page, not an empty set.
2. Free licences, several sources — nothing to license per image, and no single library's taste defining every page you publish.
3. A hosted server with its own sign-in — remote, OAuth, no local process to install and no third party's API key to paste into a config file.
4. A free tier a machine can complete on its own — no screenshot review, no contract, a monthly quota and a per-minute rate limit a worker can pace itself against.

Swipe the table sideways to see every column →

We build Pexafy, so read that last row with the scepticism it deserves — and then check it: the registry census is reproducible in one script, the Getty and Shutterstock access conditions are on their own pages, and the three photo APIs' limits are in their own docs, all linked at the bottom. The claim is narrow and falsifiable: on 18 August 2026 we could not find a second service that meets all four. If you know one, we will add it to this table — the point of writing the requirements down is that anyone can run them against any product, including ours.

What a good image server does over MCP

A connector is not a REST endpoint with a new coat of paint. Two uses of this layer are written up elsewhere and are not repeated here — illustrating a draft, in the pipeline article, and keeping a long series visually coherent, in the one on volume. What belongs to the infrastructure is the thing both depend on and neither can add afterwards: three server-side decisions, made once, for every agent that ever connects.

1. Return objects an agent can reason about. Whatever the tool returns is all the model knows — it cannot scan a grid. The difference between the two shapes below is the difference between an assistant that reasons about photographs and one that relays links:

# Typical keyword wrapper: the agent has to fetch and look, or guess
[{ "url": "https://…/photo-8795398.jpeg" }, { "url": "https://…/366611.jpg" }]

# One result, as returned — trimmed here, nothing added
{
  "rank": 1,                       // the handle a human uses: "more like #1"
  "photo_id": "019e14d9-e821-…",   // what get_similar_photos takes
  "width": 6424, "height": 4283, "orientation": "landscape",
  "color_hex": "#918872", "blur_hash": "LPI#Px?aDikCGwW?M{kD-VR*s.fl",
  "source": "Pexels", "license_type": "free",
  "alt_description": "Older couple sits together on wooden bench in a park",
  "attribution": { "plain": "Photo by Anastasia Shuraeva on Pexels", "html": "…" },
  "urls": { "thumb": "…", "small": "…", "regular": "…", "large": "…" }
}

With the second shape an agent can say “#3 is portrait, it will break your hero slot”, drop a duplicate photographer, emit the credit line, and hand your template a width/height pair that kills layout shift — without fetching a single image. With the first, every one of those decisions is a guess or a round trip.

2. Number the results, so a person can point. Results come back ranked #1, #2, #3…, which is how anyone refers to a photo in conversation — “more like #3” — with no identifiers copied by hand. In clients that support MCP Apps the thumbnails render inline, and opening one shows metadata that was already in the tool result, at no extra call.

3. Stay read-only, and be explicit about it. Three tools, no write scope, no account mutation. One of them has no keyword equivalent anywhere: search_photos_by_image takes a reference image — a client's existing hero, a screenshot pasted into the chat — plus an optional sentence to bend it (“like this, but at night”). There is no field in a tag-matching API where that request can even be typed. And an agent that cannot change anything is an agent whose worst outcome is an empty result set rather than a support ticket.

The two surfaces, and which one you want

There are exactly two ways in, and the choice is about who is in the loop rather than about capability — the tools and the catalogue are identical on both.

The connector, when a person is. One URL, https://mcp.pexafy.com/mcp, added once in an assistant's connector settings; sign-in happens in a browser window and the client receives its own credentials, so there is no key to paste into a config file and none to rotate later. Clients that do not implement OAuth send a Pexafy API key as a bearer token against the same endpoint — which is what makes the connector usable from a cron job on a build server, where nobody is around to click “Allow”. The step-by-step wiring for Claude, ChatGPT and a fleet config file is in the pipeline article; it has not changed.

The HTTP API, when no one is. A batch job filling a pool for 800 topic clusters should not pretend to be a chat client: same account, same key, plain HTTP, 100 results per call. The pacing rules that go with it — the rate-limit headers, why a per-minute 429 is worth retrying and a monthly-quota 429 is not, what a worker costs at volume — belong to the article on illustrating at scale, which is where that worker is written out in full.

The point worth keeping here is narrower, and it is the one the market study turns on: both surfaces are reachable by a program on the day it is written, with no queue and no contract.

What this does not solve

An honest list, because each of these has cost someone a rollback.

• The licence still governs the picture. Attribution is not required by the Pexafy API terms and every result ships an attribution string ready to render — but the licence attached by the originating library applies to your use of that photo. Render the credit automatically; it is cheaper than auditing later.
• Free-licence catalogues are not editorial archives. No news, no sport, no recognisable brands or public figures on demand. If your page needs a photograph of a specific event or person, that is what the licensed incumbents are for, and their connectors now exist.
• It is retrieval, and only retrieval. Diagrams, charts and screenshots are rendered from code, not searched; deduplication across your pages is a column in your database, not a parameter on the endpoint; and no photograph makes a thin page rank — Google's spam policies treat scaled content abuse the same whether or not the pages are illustrated.6 Those three are pipeline problems with pipeline answers, worked out in the pipeline article and the one on volume.

Where to start

1. Re-run the benchmark. Twenty lines, no key. Whatever image API you use today, send it three of your own agent's briefs verbatim and count the empty result sets.
2. Attach one connector to the assistant you already use and ask it, in a sentence, for a photo you actually need this week. That is the whole evaluation.
3. Move the briefs into the agent — one scene per slot, written from the draft, never the title.
4. Add the photo_id column before you publish anything at volume.
5. Run it on the free tier — 5,000 searches a month at 20 a minute until a build window, not a bill, forces you up.

References & footnotes