Claude connectors and MCP are not the same thing. MCP is the open standard, and a connector is an MCP server Claude has been authorized to call for you. This guide covers what the protocol specifies, how remote and local servers differ, which plans get connectors, and what the 816 listings in Anthropic’s Connectors Directory mean for brand discovery inside AI assistants.
In This Article
Share On:
Claude connectors and the Model Context Protocol get talked about as if they are the same thing. They are not. MCP is an open standard for connecting AI applications to external systems. A connector is what Claude calls an MCP server once you have added it, signed in and switched it on.
That one distinction explains nearly everything else: which plans get what, why some integrations only run on the desktop app, and why a listing in Anthropic’s directory is reviewed while a URL you paste in yourself is not. Here is how Claude connectors work in September 2026, what the protocol underneath them specifies, and what a public connector directory means for brands that want to be found inside an AI assistant.
The protocol is maintained in the open. Its documentation describes it as an open-source standard for connecting AI applications to external systems, and compares it to a USB-C port for AI: one plug shape, any number of devices on either side. The current specification revision is dated 2026-07-28.
Nothing in that specification is about Claude. It defines how any host application and any tool provider talk to each other, which is exactly why it spread.
A connector is an MCP server Claude can reach
Anthropic’s developer docs are direct about this: directory connectors and custom connectors run on the same MCP infrastructure. The runtime, transport, authentication and tool-calling paths are identical. What separates them is review, discoverability and distribution, not technology.
So when someone asks whether Claude connectors are MCP, the answer is yes. A connector is an MCP server with an account attached, a permission grant and a place in Claude’s interface.
Why the words get swapped
In the Claude app you never see the phrase MCP server. You see Connectors, a list of named cards with logos. Developers building the integration see the opposite view: a server, a manifest of tools, an OAuth flow. Same object, different vocabulary at each end.
What Is the Model Context Protocol?
MCP is an open standard that lets an AI application discover what a tool can do, then call it. It defines a data layer built on JSON-RPC 2.0, a transport layer that carries the messages, and a set of primitives that a server exposes to any client speaking the protocol.
Hosts, clients and servers
The architecture uses 3 roles. The host is the AI application that coordinates everything, such as Claude Desktop or Claude Code. The host creates one client per server, and each client keeps a dedicated connection. The server is, in the spec’s words, a program that provides context to MCP clients.
A server can run anywhere. That single design choice is what later splits connectors into local and remote, and it decides which Claude surfaces can use them.
Tools, resources and prompts
Servers expose 3 primitives:
Tools: executable functions the model can invoke, such as a file operation, an API call or a database query.
Resources: data sources that supply context, such as file contents or records.
Prompts: reusable templates that structure an interaction.
Clients can expose their own primitives back: sampling, so a server can ask the host for a model completion, elicitation, so it can ask the user a question, and logging. Discovery is dynamic. A client calls a list method to find out what exists, then calls a tool by name.
Who else supports it
This is not a Claude-only standard. The protocol documentation lists Claude and ChatGPT among the AI assistants that support MCP, alongside development tools including Visual Studio Code and Cursor. Build the server once and it works in more than one place, which is the whole argument for a protocol over a plug-in format.
Remote vs Local MCP Servers, and What Claude Connects To
Most MCP servers are remote. A remote server is hosted on the internet and uses the Streamable HTTP transport, so it works on every Claude surface. A local server runs on your own machine over stdio, and Anthropic is blunt about the limit: a local MCP works only in Claude Desktop and Claude Code.
Ways to connect a tool to Claude in 2026, and where each one runs
Connection type
What it is
Where it runs
Works in
Who adds it
Directory connector
A remote MCP server reviewed by Anthropic and listed in the Connectors Directory
On the provider’s servers, over Streamable HTTP
Claude.ai, Cowork, Desktop, mobile, Claude Code
Any user whose plan and organization allow it
Custom connector
A remote MCP server you add yourself by pasting its URL, with no Anthropic review
On the provider’s servers, reached from Anthropic’s cloud
The same surfaces as a directory connector
Free to Enterprise users, Owners only on Team and Enterprise
If the integration genuinely needs to touch files or apps on the machine, the packaging is different. Local MCP servers are bundled as MCP Bundles and distributed as desktop extensions for Claude Desktop, and they go through a separate submission form rather than the directory portal.
What Claude supports on the wire
Claude accepts Streamable HTTP and the legacy HTTP plus SSE transport, with the legacy option being deprecated. Tool results on Claude.ai and Desktop are capped near 150,000 characters, and a single tool call times out after 240 seconds. Those are useful numbers when you are deciding how much work one tool should do.
Individual users add a custom connector from Customize, then Connectors, then Add custom connector, and paste the server URL. Anthropic attaches a warning worth repeating: custom connectors let you connect Claude to arbitrary services that have not been verified by Anthropic. Connect to servers you trust, read the permissions, and stay alert to prompt injection.
Team and Enterprise
On Team and Enterprise plans, only Owners add a connector for the organization. Members then connect their own accounts to it. Owners can set what each connected service is allowed to do, with actions marked Always allow, Needs approval or Blocked. Team members without that permission see a Request button on a directory listing, which sends it to an admin.
Where they work
The same catalog serves Claude.ai, Cowork, Claude Desktop, mobile and Claude Code. Skills, connectors and plugins now sit in one directory under Customize, with a separate tab for anything shared across your organization.
What Is Inside the Connectors Directory
The public Connectors Directory listed 816 connectors when we checked it in September 2026. You can filter by category and by type (web, desktop extension or interactive) and sort by recommended, most popular, trending, newest or alphabetically. Categories run from developer tools and productivity to healthcare, legal and financial services.
Verified and community listings
The directory carries both labels. Verified connectors have been tested by Anthropic for quality and compatibility and met the Software Directory Policy at the time of review, though Anthropic is careful to say verification is not a security audit. Community connectors are screened before listing but not reviewed in depth.
The label describes the review, not the power. Once connected, a community connector has the same capabilities and access as any connector you grant.
The tool list is longer than any blog post can hold
Older roundups of Claude integrations, this one included, used to publish a fixed table of supported apps. With 816 listings and a submission portal that is always open, that table is out of date the day it is written. Browse the live directory instead, then read the individual guides for the tools you actually use.
What Connector Directories Mean for Brand Discovery
A directory listing is a new shelf, and it behaves like an app store. Directory connectors are eligible for Suggested Connectors, which are in-chat recommendations Claude makes when your tool is relevant to the task. Every listing is included automatically, and ranking is usage-based. A connector added by URL gets none of that.
What a listing actually buys you
A listed connector can be browsed and searched inside the product, shows up as a named card with your logo rather than the word Custom, and gets a permanent listing URL you can put in your docs, your onboarding email or a Connect to Claude button. A custom connector has no in-product discovery at all.
For a SaaS product, that is a distribution question, not a content question. Your tool becomes reachable at the moment of intent, inside the assistant, without the user opening your app first.
Why it is not a free traffic channel
Usage-based ranking rewards products that already have install bases. With 816 listings and more arriving through an always-open portal, the middle of that list is quiet. Being present is table stakes, and it does not answer the question that comes earlier in the buying journey: when someone asks an AI assistant which tool to use, does your brand get named?
Getting listed is a product commitment. Submission runs through a portal inside Claude.ai organization settings, so you need a Team or Enterprise organization and directory management access. Individual plans cannot submit.
Your server also has to meet Anthropic’s security standards, use OAuth 2.0 for authenticated services, and annotate every tool with a title and either a read-only or a destructive hint. Local connectors must publish a privacy policy, and a missing one is an immediate rejection.
The listing itself is tightly framed: a name up to 100 characters, a tagline up to 55, a description up to 2,000, and 1 to 5 categories. That tagline is the entire positioning exercise, and you write it, not Anthropic.
Publishing elsewhere does not put you in Claude
Worth knowing before someone on your team ships to the wrong place. The Anthropic directory is independent of the open MCP Registry and the public servers repository. Publishing to those does not surface your server in Claude. Only a directory submission does.
Claude Connectors and MCP: FAQs
What is the difference between Claude connectors and MCP?
MCP is the open standard and a connector is Claude’s version of it. The Model Context Protocol defines how an AI application discovers what a tool can do and then calls it. A connector is an MCP server that Claude has been pointed at, authorized against and allowed to use. Anthropic’s docs confirm directory and custom connectors run on the same MCP infrastructure, so the difference is review and discovery, not technology.
Which Claude plans include connectors?
All of them. Anthropic’s help centre says web connectors are available for all users on Claude, Cowork, Claude Desktop and Claude Mobile, and that custom connectors using remote MCP work on Free, Pro, Max, Team and Enterprise plans. Free accounts are limited to 1 custom connector. On Team and Enterprise plans, only Owners add a connector for the organization, after which members connect their own accounts to it.
What is the difference between a remote and a local MCP server?
A remote MCP server is hosted on the internet and uses the Streamable HTTP transport, so Claude can reach it from any surface, including web and mobile. A local server runs on your own machine over stdio and works only in Claude Desktop and Claude Code. Most MCP servers are remote. Custom connectors are reached from Anthropic’s cloud, so your server must be reachable over the public internet.
How many connectors does Claude have?
The public directory listed 816 connectors in September 2026, across categories including productivity, developer tools, sales and marketing, financial services, healthcare and legal. You can filter by category and by type, meaning web, desktop extension or interactive, and sort by recommended, most popular, trending, newest or alphabetically. The same catalog serves Claude.ai, Cowork, Desktop, mobile and Claude Code.
How do I get my product listed in the Connectors Directory?
Submissions run through a portal inside Claude.ai organization settings, so you need a Team or Enterprise organization and directory management access. Your remote MCP server must meet Anthropic’s security standards, use OAuth 2.0 for authenticated services, and give every tool a title and either a read-only or a destructive hint. Local connectors need a privacy policy, and a missing one is rejected immediately.
Do Claude connectors help my brand get discovered?
A listing is a discovery surface rather than a traffic channel. Directory connectors are eligible for Suggested Connectors, the in-chat recommendations Claude makes when a tool fits the task, and every listing is included automatically. Ranking is usage-based, similar to other app stores, so products with install bases sit at the top. For most brands, being named in the answer still matters more.
Your Next Move
Pick the layer you actually need. If you want Claude to work across the tools your team already pays for, start in the directory and connect what is there. If you build software, the question is sharper: can a model use your product without a human clicking through your interface?
At upGrowth, we help growth and product teams answer that question in both directions, making the product usable by AI agents and making the brand visible in the answers those agents give.
The Model Context Protocol (MCP) is the underlying open standard, while a connector is Anthropic's productized implementation of an MCP server within the Claude interface. This means MCP defines the universal rules for how any AI application can talk to an external tool, whereas a connector is what you see and interact with after an MCP server has been connected and authorized in your Claude account. This separation explains why the core technology is identical for both public directory connectors and private custom ones, with the only differences being review, discoverability, and distribution.
This distinction directly impacts how tools are developed and deployed. For instance, a developer builds one MCP server that can theoretically work with multiple AI hosts, not just Claude. For a user, knowing this clarifies why some connectors are vetted and listed in a directory while others, added via a URL, are not. The protocol itself is open-source, with the current specification dated 2026-07-28, ensuring a stable, documented foundation for anyone to build on. Understanding this split between the protocol and the product is the first step to mastering how to extend Claude's capabilities safely and effectively.
The Model Context Protocol's architecture defines three distinct roles to ensure a clear separation of concerns: the host, the client, and the server. The host is the primary AI application, such as Claude Desktop, that orchestrates the interaction. For each external tool, the host creates a dedicated client that maintains the connection. The server is the external program providing context, tools, or data, which can run anywhere, from a public cloud to a local machine.
This modular design is the key to MCP's flexibility and broad adoption. The specification allows a server to expose three main primitives that the host's client can dynamically discover and use:
Tools: These are executable functions the AI can invoke, like an API call or a database query.
Resources: These are data sources that provide context, such as the contents of a file.
Prompts: These are reusable templates for structuring interactions.
This architecture explains why MCP is not a Claude-only standard and is supported by applications like Visual Studio Code. By defining these clear roles and primitives, the protocol provides a common language for any AI to connect with any tool, a concept the documentation compares to a USB-C port for AI.
A remote MCP server is hosted on the public internet, whereas a local MCP server runs directly on your own machine. This fundamental difference in location dictates which Claude surfaces can use the integration. Remote servers, which use the Streamable HTTP transport, are accessible from all Claude platforms, including claude.ai, mobile apps, and Claude Desktop. In contrast, local servers are restricted; Anthropic is explicit that a local MCP server running over stdio only works within Claude Desktop and Claude Code.
Choosing between remote and local is a strategic decision about reach versus data access. If your goal is broad distribution and availability across all of Claude's platforms, a remote server is the only option. However, if your tool must interact with local files or applications on a user's machine, a local server packaged as a desktop extension is necessary. Developers must also remember that custom connectors added via URL connect from Anthropic's cloud, meaning even for private use, the server must be publicly reachable, a detail that often trips up teams building internal tools behind a firewall.
The presence of 816 connectors in the directory indicates a rapidly maturing ecosystem where Anthropic is balancing openness with curation. The directory allows users to filter by categories like 'healthcare' and 'developer tools' and sort by metrics such as 'most popular' and 'trending,' mimicking the structure of established app stores. This signals a strategic move to make third-party tools a core, discoverable part of the Claude user experience.
The distinction between 'Verified' and 'Community' labels reveals a tiered approach to trust and quality control. 'Verified' connectors have undergone a quality and compatibility review by Anthropic, meeting its Software Directory Policy. 'Community' connectors are screened but not reviewed in depth. This system allows Anthropic to grow the directory quickly while guiding users toward integrations that have received more scrutiny. However, it's crucial to understand that verification is not a security audit, and a community connector has the same technical capabilities as a verified one once permission is granted, placing the final responsibility on the user.
Cross-platform support for the Model Context Protocol (MCP) by assistants like Claude and ChatGPT, alongside developer tools such as Visual Studio Code, highlights the strategic advantage of a 'build once, run anywhere' approach. For developers, this means the significant effort invested in creating an MCP server is not locked into a single ecosystem. An integration built for one AI assistant can be deployed in another with minimal changes, drastically reducing development overhead and expanding the potential user base.
This interoperability is the core argument for an open protocol over a closed plug-in system. While a proprietary format might offer deeper integration with one specific platform, MCP provides broader market access. This makes it a more attractive proposition for SaaS companies and independent developers who want their tools to be available wherever users are working. As the number of AI applications grows, the value of this portability increases, positioning MCP as a key piece of infrastructure for the emerging agentic software landscape.
Getting a product listed in the Claude Connectors Directory is a formal process that requires both organizational access and technical compliance. It is not available to individual users. The process begins with having a Team or Enterprise plan, as submissions are handled exclusively through a portal within the organization settings, and requires a user to have directory management access.
Beyond simple access, your MCP server must meet a specific set of standards to be accepted. Here is the implementation path:
Technical Setup: Your remote MCP server must use OAuth 2.0 for any authenticated services. This is a non-negotiable security requirement.
Tool Annotation: Every tool your server exposes must be annotated with a title and a hint indicating if it is 'read-only' or 'destructive'.
Policy Compliance: For local connectors packaged as desktop extensions, a publicly accessible privacy policy is mandatory, and its absence leads to immediate rejection.
Listing Content: You must provide tightly framed marketing copy, including a name (up to 100 characters), a tagline (up to 55 characters), and a description (up to 2,000 characters).
This entire process underscores that a directory listing is a product commitment, not just a simple registration, demanding adherence to Anthropic's ecosystem rules.
Users on a Claude Free plan are permitted to add one custom connector by providing its server URL directly. This allows you to integrate a tool that hasn't been listed in the public directory. The process is straightforward and managed within the application's settings, designed to give users flexibility in extending Claude's functionality.
To add your connector, you must follow these steps inside the Claude interface: Navigate from Customize, to Connectors, and then select Add custom connector. At this point, you will be prompted to paste the URL of the MCP server you wish to connect. During this process, Anthropic attaches a significant warning that you should not ignore: 'custom connectors let you connect Claude to arbitrary services that have not been verified by Anthropic.' This message emphasizes that the responsibility for vetting the service's trustworthiness lies entirely with you. Always connect only to servers you trust, carefully review the permissions requested, and remain vigilant against potential prompt injection attacks, as this path bypasses all of Anthropic's official review processes.
The usage-based ranking for 'Suggested Connectors' creates a feedback loop where popular tools become more visible, further cementing their market position. As the Claude Connectors Directory expands, this mechanism will likely lead to a power-law distribution, with a small number of top-ranked connectors capturing a large share of user discovery and engagement. New entrants will find it increasingly difficult to gain organic visibility purely through a directory listing.
This forces brands to treat the directory as a distribution channel, not a free marketing one. Your strategy must evolve beyond simply being listed. A key adjustment is to focus on driving initial adoption through your own channels, such as onboarding emails or documentation with a 'Connect to Claude' button, to generate the usage signals that feed the ranking algorithm. Furthermore, you must also solve the earlier-stage discovery problem: ensuring your brand is mentioned when a user asks an AI assistant for a tool recommendation. This requires a parallel content strategy focused on making your product's value proposition easily citable by AI models, a challenge entirely separate from the technical integration.
The integration of the Model Context Protocol (MCP) into developer tools like Visual Studio Code and Cursor signals a shift where AI assistants are becoming active participants in the development lifecycle, not just passive chat interfaces. This moves beyond simple code generation. By using MCP, an AI can now directly and safely invoke development tools, access local files, and interact with the developer's environment in a structured way.
This trend points toward a future of deeply embedded, agentic AI partners in software engineering. For developers, it means AI assistants can perform complex, multi-step tasks like running tests, refactoring code across multiple files, or querying a local database for context, all orchestrated through a standard protocol. The open nature of MCP ensures that developers are not locked into a single AI provider, like Anthropic, and can build or adopt tools that work across their preferred environments. This deeper integration promises to accelerate development workflows and change the very nature of how programmers interact with their tools.
A frequent mistake teams make is assuming that a 'custom connector' added on claude.ai can access an MCP server running behind their corporate firewall. This assumption leads to failed connections because it misunderstands how Anthropic's infrastructure works. The official help center documentation is clear: custom connectors connect to your MCP server from Anthropic's cloud, not from the user's local device.
The solution requires making your internal tool's MCP server accessible to the public internet. Your server cannot be on a private network or localhost. If the tool genuinely needs to access local machine resources, the correct approach is not a custom connector. Instead, the integration must be packaged as a desktop extension for Claude Desktop. This uses a local MCP server that communicates over stdio, a completely different connection path. Understanding this distinction between cloud-initiated remote connections and locally-initiated desktop connections is critical for successfully integrating internal enterprise tools with Claude.
Developers often assume that because the Model Context Protocol (MCP) is an open standard, any public registry associated with it, like the open MCP Registry or public servers repository, will automatically feed into major applications like Claude. This is incorrect. These registries are for the open-source community and general protocol discoverability, but they are completely independent of Anthropic's product ecosystem.
The exclusive pathway to get a tool listed and discoverable within the Claude application is through a formal directory submission. Anthropic maintains its own private, curated Connectors Directory. To be included, you must have a Team or Enterprise plan and use the dedicated submission portal found within your organization's settings on claude.ai. This process involves meeting specific security standards, using OAuth 2.0, and passing a review. Publishing anywhere else will not make your server visible within the Claude interface, a crucial distinction for any brand seeking in-product discovery.
It is a mistake to equate the 'verified' label with comprehensive security and 'community' with being harmless. The label describes the depth of Anthropic's review process, not the technical capabilities of the connector. A 'verified' connector has been checked for quality and compatibility, but this is explicitly not a full security audit. Once connected, a community connector has the exact same potential to access your data and perform actions as a verified one.
You should operate on the principle of least privilege when granting permissions to any connector, regardless of its label. Before authorizing any service, carefully scrutinize the permissions it requests and consider whether it truly needs that level of access to perform its function. For organizations on Team and Enterprise plans, Owners can enforce this by setting actions for each service to 'Always allow,' 'Needs approval,' or 'Blocked.' For all users, the core security mindset remains the same: the 'verified' label offers a degree of confidence, but the ultimate responsibility for protecting your information rests with you.
Amol has helped catalyse business growth with his strategic & data-driven methodologies. With a decade of experience in the field of marketing, he has donned multiple hats, from channel optimization, data analytics and creative brand positioning to growth engineering and sales.
