700 MCP servers, mapped: where AI agents are really going

An AI agent is only as useful as the things it can touch. The Model Context Protocol — MCP — is the plug through which agents touch them: a common way for a model to call a database, a search engine, a browser or a code host. We categorised all 700 MCP servers tracked in the olud.ai hub to see where that building effort is actually going. The result is not a picture of a mature market. It is a picture of a rush.

The biggest category is the absence of one

Start with the bar that should not exist. “Other” — the bucket for servers that fit no established label — is the largest category on the map: 255 servers, 36.4% of everything we track. In a settled ecosystem, the miscellaneous shelf is the small one at the back of the shop. Here it is the front window.

The runner-up makes the same point from the opposite side. Aggregators & Toolkits — servers whose purpose is to route, bundle or manage other servers — count 157 entries, 22.4% of the map. This is infrastructure for infrastructure, a category that only grows this large when a protocol is young enough that its users need help holding the pieces together. Part of what aggregators sell is choice across models and providers whose prices differ sharply — the same spread we measured in the open vs closed LLM price gap.

Put the unclassifiable next to the meta-tools and they dominate the chart. That is the leading fact about the MCP ecosystem right now: it is being built faster than it can be named. Standards produce stable categories; land rushes produce a very large pile marked “miscellaneous” and a brisk trade in shovels.

Memory before hands

Filter out the noise and the picture sharpens. The largest single-purpose category on the map is Memory & Knowledge: 68 servers, 9.7% of the total. Knowledge bases, note stores, retrieval layers — anything that lets an agent carry facts from one session into the next.

That ordering is worth pausing on. The public image of an AI agent is a thing that does: books the flight, files the pull request, answers the ticket. The building activity says otherwise. Before agents get hands, they are getting memories. An assistant that forgets everything when the window closes is a demo; the ecosystem is spending its energy on the unglamorous layer that turns demos into colleagues.

Memory is also the layer people most want to keep close. A retrieval store holds exactly the data you would rather not hand to a third party, which is why the category pairs naturally with running models on your own hardware — the terrain of our local AI guide and of what runs on your GPU.

The working middle: read, query, commit

Below memory sits the everyday toolkit of a working agent. Search & Web counts 66 servers (9.4%) — an agent that cannot look things up is blind. Databases follow at 53 (7.6%), GitHub & Git at 37 (5.3%), and Browsers & Automation at 32 (4.6%).

One detail in that middle is telling: there are more servers for searching the web than for driving a browser. Given the choice, builders wire agents into structured APIs rather than teach them to click. A browser is what an agent falls back on when nothing better exists — and the map suggests something better usually exists.

The tail is the honest part

The bottom of the chart is where the map stops flattering the ecosystem. Cloud & DevOps counts 18 servers (2.6%). Messaging counts 9 (1.3%). Filesystem — the plainest capability imaginable — counts 5, or 0.7% of the map.

Small numbers here mean opposite things. Filesystem is tiny because the job is solved elsewhere: every serious agent harness ships file access natively, so almost nobody needs it as a separate server. Messaging is tiny for a more revealing reason — a server that lets an agent send messages on your behalf is a server that lets an agent embarrass you at scale, and builders are visibly not lining up to offer that. Cloud & DevOps sits low on similar grounds: production infrastructure is where tolerance for a confused agent runs out.

A census taken mid-stampede

Everything above describes a protocol in its land-rush phase: the biggest shelf is labelled miscellaneous, and much of the rest is tools for coping with the other tools. None of this is a criticism — it is what the early years of every platform look like — but it does come with a prediction: consolidation. Rushes end with abandoned claims, and we have measured what that looks like elsewhere, in how much of open-source AI is already abandoned: projects go quiet long before their download counts do.

Expect this map's Other bucket to shrink, its aggregators to merge, and a meaningful share of today's 700 servers to stop being maintained. Our project health tracker exists for exactly that moment — checking whether the server you are about to depend on still has anyone behind it.

Full disclosure: this hub is on the map

One honesty note before the method. olud.ai is not a neutral cartographer here: the hub itself speaks MCP. An agent can query our catalogue — projects, models, alternatives — through the endpoint documented on our MCP page, which makes us a resident of the very map we are drawing, somewhere in the aggregator district. We think living on the map is the best reason to draw it carefully, and how the data behind it is built and refreshed is public on the methodology page.

Method

The counts come from mcp-map.json, the category index behind the olud.ai MCP hub: 700 servers, each assigned to the single category that best matches its primary function. A server that plugs into a database and also does a little search is filed under Databases — primary function wins. When no established label fit, we filed the server under Other rather than invent a category per server; the size of that bucket is a finding about the ecosystem, not a shortcut in the data. Percentages are shares of the 700 total. The snapshot is 19 August 2026; the underlying catalogue is refreshed nightly.

Keep reading

• How much of open-source AI is already abandoned — what maintenance data says about the projects everyone still recommends.
• The open vs closed LLM price gap — what open-weight models actually cost next to the closed ones.
• What runs on your GPU — which models fit on the hardware people actually own.

Data: olud.ai catalogue, refreshed nightly — method on /methodology.html.