MCP in Practice | Towards Data Science

Written along with Ilan Strauss, Isobel Moure, and Tim O’Reilly as a part of the AI Disclosures Project. Initially printed on our weblog: Asimov’s Addendum.

1. of MCP

Anthropic’s Model Context Protocol (MCP) was launched in November 2024 as a strategy to make instruments and platforms model-agnostic. MCP works by defining servers and shoppers. MCP servers are native or distant finish factors the place instruments and assets are outlined. For instance, GitHub launched an MCP server that permits LLMs to each learn from and write to GitHub. MCP shoppers are the connection from an AI software to MCP servers—they permit an LLM to work together with context and instruments from totally different servers. An instance of an MCP consumer is Claude Desktop, which permits the Claude fashions to work together with hundreds of MCP servers.

In a comparatively brief time, MCP has change into the spine of tons of of AI pipelines and functions. Main gamers like Anthropic and OpenAI have constructed it into their merchandise. Developer instruments reminiscent of Cursor (a coding-focused textual content editor or IDE) and productiveness apps like Raycast additionally use MCP. Moreover, hundreds of developers use it to combine AI fashions and entry exterior instruments and information with out having to construct a whole ecosystem from scratch.

In earlier work printed with AI Frontiers, we argued that MCP can act as an ideal unbundler of “context”—the info that helps AI functions present extra related solutions to customers. In doing so, it may well assist decentralize AI markets. We argued that, for MCP to really obtain its objectives, it requires help from:

1. Open APIs: In order that MCP functions can entry third-party instruments for agentic use (write actions) and context (learn)
2. Fluid reminiscence: Interoperable LLM reminiscence requirements, accessed through MCP-like open protocols, in order that the reminiscence context accrued at OpenAI and different main builders doesn’t get caught there, stopping downstream innovation

We broaden upon these two factors in a recent policy note, for these trying to dig deeper.

Extra usually, we argue that protocols, like MCP, are literally foundational “rules of the road” for AI markets, whereby open disclosure and communication requirements are constructed into the community itself, somewhat than imposed after the actual fact by regulators. Protocols are basically market-shaping units, architecting markets via the permissions, guidelines, and interoperability of the community itself. They will have a big effect on how the industrial markets constructed on high of them perform too.

1.1 However how is the MCP ecosystem evolving?

But we don’t have a transparent thought of the form of the MCP ecosystem right now. What are the most typical use circumstances of MCP? What kind of entry is being given by MCP servers and utilized by MCP shoppers? Is the info accessed through MCP “read-only” for context, or does it enable brokers to “write” and work together with it—for instance, by modifying information or sending emails?

To start answering these questions, we have a look at the instruments and context which AI brokers use through MCP servers. This offers us a clue about what’s being constructed and what’s getting consideration. On this article, we don’t analyze MCP shoppers—the functions that use MCP servers. We as an alternative restrict our evaluation to what MCP servers are making out there for constructing.

We assembled a big dataset of MCP servers (n = 2,874), scraped from Pulse. We then enriched it with GitHub star-count information on every server. On GitHub, stars are just like Fb “likes,” and developers use them to indicate appreciation, bookmark tasks, or point out utilization.

In follow, whereas there have been loads of MCP servers, we discovered that the highest few garnered many of the consideration and, possible by extension, many of the use. Simply the highest 10 servers had practically half of all GitHub stars given to MCP servers.

A few of our takeaways are:

1. MCP utilization seems to be pretty concentrated. Because of this, if left unchecked, a small variety of servers and (by extension) APIs might have outsize management over the MCP ecosystem being created.
2. MCP use (instruments and information being accessed) is dominated by simply three classes: Database & Search (RAG), Pc & Internet Automation, and Software program Engineering. Collectively, they obtained practically three-quarters (72.6%) of all stars on GitHub (which we proxy for utilization).
3. Most MCP servers help each learn (entry context) and write (change context) operations, displaying that builders need their brokers to have the ability to act on context, not simply eat it.

2. Findings

To begin with, we analyzed the MCP ecosystem for focus danger.

2.1 MCP server use is concentrated

We discovered that MCP utilization is concentrated amongst a number of key MCP servers, judged by the variety of GitHub stars every repo obtained.

Regardless of there being hundreds of MCP servers, the highest 10 servers make up practically half (45.7%) of all GitHub stars given to MCP servers (pie chart beneath) and the highest 10% of servers make up 88.3% of all GitHub stars (not proven).

Because of this the vast majority of real-world MCP customers are possible counting on the identical few providers made out there through a handful of APIs. This focus possible stems from community results and sensible utility: All builders gravitate towards servers that remedy common issues like internet shopping, database entry, and integration with broadly used platforms like GitHub, Figma, and Blender. This focus sample appears typical of developer-tool ecosystems. Just a few well-executed, broadly relevant options are likely to dominate. In the meantime, extra specialised instruments occupy smaller niches.

2.2 The highest 10 MCP servers actually matter

Subsequent, the highest 10 MCP servers are proven within the desk beneath, together with their star depend and what they do.

Among the many high 10 MCP servers, GitHub, Repomix, Context7, and Framelink are constructed to help with software program improvement: Context7 and Repomix by gathering context, GitHub by permitting brokers to work together with tasks, and Framelink by passing on the design specs from Figma on to the mannequin. The Blender server permits brokers to create 3D fashions of something, utilizing the favored open supply Blender software. Lastly, Activepieces and MindsDB join the agent to a number of APIs with one standardized interface: in MindsDB’s case, primarily to learn information from databases, and in Activepieces to automate providers.

The dominance of agentic shopping, within the type of Browser Use (61,000 stars) and Playwright MCP (18,425 stars), stands out. This displays the basic want for AI techniques to work together with internet content material. These instruments enable AI to navigate web sites, click on buttons, fill out varieties, and extract information identical to a human would. Agentic shopping has surged, despite the fact that it’s far much less token-efficient than calling an API. Shopping brokers usually have to wade via a number of pages of boilerplate to extract slivers of knowledge a single API request might return. As a result of many providers lack usable APIs or tightly gate them, browser-based brokers are sometimes the only—generally the one—strategy to combine, underscoring the boundaries of right now’s APIs.

Among the high servers are unofficial. Each the Framelink and Blender MCP are servers that work together with only a single software, however they’re each “unofficial” merchandise. Because of this they aren’t formally endorsed by the builders of the applying they’re integrating with—those that personal the underlying service or API (e.g., GitHub, Slack, Google). As an alternative, they’re constructed by impartial builders who create a bridge between an AI consumer and a service—usually by reverse-engineering APIs, wrapping unofficial SDKs, or utilizing browser automation to imitate consumer interactions.

It’s wholesome that third-party builders can construct their very own MCP servers, since this openness encourages innovation. But it surely additionally introduces an middleman layer between the consumer and the API, which brings dangers round belief, verification, and even potential abuse. With open supply native servers, the code is clear and will be vetted. Against this, distant third-party servers are more durable to audit, since customers should belief code they will’t simply examine.

At a deeper degree, the repos that presently dominate MCP servers spotlight three encouraging info concerning the MCP ecosystem:

1. First, a number of outstanding MCP servers help a number of third-party providers for his or her performance. MindsDB and Activepieces function gateways to a number of (usually competing) service suppliers via a single server. MindsDB permits builders to question totally different databases like PostgreSQL, MongoDB, and MySQL via a single interface, whereas Taskmaster permits the agent to delegate duties to a variety of AI fashions from OpenAI, Anthropic, and Google, all with out altering servers.
2. Second, agentic shopping MCP servers are getting used to get round doubtlessly restrictive APIs. As famous above, Browser Use and Playwright entry web providers via an online browser, serving to to bypass API restrictions, however they as an alternative run up towards anti-bot protections. This circumvents the constraints that APIs can impose on what builders are in a position to construct.
3. Third, some MCP servers do their processing on the developer’s pc (regionally), making them much less depending on a vendor sustaining API entry. Some MCP servers examined right here can run solely on an area pc with out sending information to the cloud—that means that no gatekeeper has the facility to chop you off. Of the ten MCP servers examined above, solely Framelink, Context7, and GitHub depend on only a single cloud-only API dependency that may’t be run regionally end-to-end in your machine. Blender and Repomix are fully open supply and don’t require any web entry to work, whereas MindsDB, Browser Use, and Activepieces have native open supply implementations.

2.3 The three classes that dominate MCP use

Subsequent, we grouped MCP servers into totally different classes primarily based on their performance.

Once we analyzed what varieties of servers are hottest, we discovered that three dominated: Pc & Internet Automation (24.8%), Software program Engineering (24.7%), and Database & Search (23.1%).

Widespread use of Software program Engineering (24.7%) MCP servers aligns with Anthropic’s economic index, which discovered that an outsize portion of AI interactions had been associated to software program improvement.

The recognition of each Pc & Internet Automation (24.8%) and Database & Search (23.1%) additionally is smart. Earlier than the appearance of MCP, internet scraping and database search had been extremely built-in functions throughout platforms like ChatGPT, Perplexity, and Gemini. With MCP, nonetheless, customers can now entry that very same search performance and join their brokers to any database with minimal effort. In different phrases, MCP’s unbundling impact is extremely seen right here.

2.4 Brokers work together with their environments

Lastly, we analyzed the capabilities of those servers: Are they permitting AI functions simply to entry information and instruments (learn), or as an alternative do agentic operations with them (write)?

Throughout all however two of the MCP server classes checked out, the most well-liked MCP servers supported each studying (entry context) and writing (agentic) operations—proven in turquoise. The prevalence of servers with mixed learn and write entry means that brokers should not being constructed simply to reply questions primarily based on information but additionally to take motion and work together with providers on a consumer’s behalf.

The 2 exceptions are Database & Search (RAG) and Finance MCP servers, during which read-only entry is a standard permission given. That is possible as a result of information integrity is essential to making sure reliability.

3. The Significance of A number of Entry Factors

Just a few implications of our evaluation will be drawn out at this preliminary stage.

First, concentrated MCP server use compounds the dangers of API entry being restricted. As we mentioned in “Protocols and Power,” MCP stays constrained by “what a specific service (reminiscent of GitHub or Slack) occurs to show via its API.” Just a few highly effective digital service suppliers have the facility to close down entry to their servers.

One vital hedge towards API gatekeeping is that lots of the high servers strive to not depend on a single presentr. As well as, the next two safeguards are related:

• They provide native processing of knowledge on a consumer’s machine each time doable, as an alternative of sending the info for processing to a third-party server. Native processing ensures that performance can’t be restricted.
• If working a service regionally will not be doable (e.g., e mail or internet search), the server ought to nonetheless help a number of avenues of getting on the wanted context via competing APIs. For instance, MindsDB features as a gateway to a number of information sources, so as an alternative of counting on only one database to learn and write information, it goes to nice lengths to help a number of databases in a single unified interface, primarily making the backend instruments interchangeable.

Second, our evaluation factors to the truth that present restrictive API entry insurance policies should not sustainable. Internet scraping and bots, accessed through MCP servers, are in all probability getting used (at the very least partly) to avoid overly restrictive API entry, complicating the increasingly common follow of banning bots. Even OpenAI is coloring exterior the API strains, utilizing a third-party service to entry Google Search’s outcomes via internet scraping, thereby circumventing its restrictive API.

Increasing structured API entry in a significant method is significant. This ensures that reputable AI automation runs via secure, documented finish factors. In any other case, builders resort to brittle browser automation the place privateness and authorization haven’t been correctly addressed. Regulatory steerage could push the market on this path, as with open banking within the US.

Lastly, encouraging better transparency and disclosure might assist establish the place the bottlenecks within the MCP ecosystem are.

• Builders working well-liked MCP servers (above a sure utilization threshold) or offering APIs utilized by high servers ought to report utilization statistics, entry denials, and rate-limiting insurance policies. This information would assist regulators establish rising bottlenecks earlier than they change into entrenched. GitHub would possibly facilitate this by encouraging these disclosures, for instance.
• Moreover, MCP servers above sure utilization thresholds ought to clearly record their dependencies on exterior APIs and what fallback choices exist if the first APIs change into unavailable. This isn’t solely useful in figuring out the market construction, but additionally important info for safety and robustness for downstream functions.

The aim is to not eradicate all focus within the community however to make sure that the MCP ecosystem stays contestable, with a number of viable paths for innovation and consumer alternative. By addressing each technical structure and market dynamics, these prompt tweaks might assist MCP obtain its potential as a democratizing power in AI improvement, somewhat than merely shifting bottlenecks from one layer to a different.

Appendix

Dataset

For this evaluation, we categorized MCP servers discovered on PulseMCP.com into one among 15 classes utilizing GPT-5 mini. We then human-reviewed and edited the highest 50 servers that make up round 70% of the whole star depend in our dataset.

The complete dataset, together with descriptions of the classes, will be discovered right here (constructed by Sruly Rosenblat):

https://huggingface.co/datasets/sruly/MCP-In-Practice

Limitations

There are a number of limitations to our preliminary analysis:

• GitHub stars aren’t a measure of obtain counts and even essentially a repo’s recognition.
• Solely the title and outline had been used when categorizing repos with the LLM.
• Categorization was topic to each human and AI errors and plenty of servers would possible match into a number of classes.
• We solely used the PulseMCP record for our dataset; different lists had totally different servers (e.g., Browser Use isn’t on mcpmarket.com).
• We excluded some repos from our evaluation, reminiscent of those who had a number of servers per code repository and people we weren’t in a position to fetch the star depend for. We additionally solely checked out servers listed on PulseMCP. Our server record is non exhaustive.

MCP Server Use Over Time