MCP Distribution · Claude
Anthropic runs the most structured of the three submission processes, and the most documented. The engineering is rarely what holds a submission up. The organization plan, the tool annotations, and the privacy policy are.
The shape of it
The Connectors Directory is a reviewed collection of MCP servers. Anyone can build and run a server for Claude without going near it; a listing is what makes the server discoverable to people who are not already your customers. Submission does not change the tools your server exposes.
Three things can be submitted: remote MCP servers, desktop extensions packaged as MCP Bundles, and MCP Apps that surface interactive UI. Only the first two are common for a company publishing its own data, and they take different routes. Remote servers go through the portal in your organization settings. Desktop extensions use a separate form.
As of August 2026, Node8's directory research counted 1,253 connectors listed in the Claude directory. The State of Claude Connectors covers what those listings actually contain, including how often published tool counts disagree with what the live endpoint returns.
Before the portal
The submission portal lives in your organization settings on Claude.ai. Individual plans do not have organization settings, so they cannot submit at all. This is the single most common reason a team gets stuck before it starts: the engineer who built the server is on a personal plan.
By default only Owners and Primary Owners can submit and manage listings. On Enterprise an Owner can delegate by creating a custom role with the Directory permission, or the broader Libraries permission that also covers plugins, connectors and skills. Team plans have no custom roles, so on Team it stays with Owners.
The portal accepts remote MCP servers only, over streamable HTTP or SSE. Local servers go a different route entirely, as desktop extensions or plugins. You also declare whether every user connects to the same URL or different users get different URLs.
Each tool needs a title and the applicable readOnlyHint or destructiveHint. The portal syncs your tools from the live server and groups them by annotation, flagging any that are missing. Fix these on the server before you submit rather than discovering them in the portal.
Supported modes are OAuth with dynamic client registration, client ID metadata documents, or a static client ID held by Anthropic. There is also a custom connection mode where users supply their own URL or credentials. Some of these need coordination with the review team rather than working out of the box.
Anthropic states that missing or incomplete privacy policies result in immediate rejection. It must cover data collection, usage and storage, third-party sharing, retention, and contact information. This is the cheapest rejection to avoid and one of the most common.
The portal
Submission happens inside Claude.ai. Progress saves in your browser as you move between steps, so you can jump back within a session without losing work. Have your documentation URL, privacy policy URL, icon and test credentials ready before you start.
Sets expectations: a listing makes the connector discoverable, it does not change the tools it exposes.
The server URL, the transport, and whether all users share one URL.
Tools, prompts and resources sync from the live server and are grouped by their annotations.
Name up to 100 characters, tagline up to 55, description up to 2,000, one to five categories, documentation and privacy URLs, support contact, icon, and the URL slug.
Primary use cases, what a user needs before connecting, and whether the connector reads, writes, or both.
Company name, website, and a contact for review updates.
Which auth mode, including the case where the server starts unauthenticated and individual tools prompt on demand.
Whether the API is yours, proxied from a partner with permission, or a third party you do not control, plus health data and sponsored content flags.
Test-account credentials and access steps detailed enough for a reviewer to run every tool end to end. You confirm you have run them all yourself.
Seven acknowledgments covering directory guidelines, first-party API usage, financial transactions, AI media generation, prompt injection, conversation data collection, and public documentation. All seven are required.
A final read-through. Quality warnings, such as very short answers, are shown here and passed to the review team with your submission.
Two of these deserve more time than they look like they need. Test & launch is where most avoidable delay comes from: a reviewer has to be able to reach your server and run every tool, so the credentials need to work against a populated account, not an empty one. And Listing contains the slug, which is permanent once published.
In practice
The pattern we see is that the server works long before the submission does. A team finishes the tools, tests them in Claude as a custom connector, and then loses weeks to things that have nothing to do with engineering: the organization is on the wrong plan, the person with Owner rights is on holiday, the privacy policy is a page on a marketing site that does not mention retention, or the tools were written without annotations because nothing enforced them during development.
Annotations are worth treating as a build requirement rather than a submission chore. They are also required by the other two platforms, in near-identical form, so doing them once at build time removes the same task from all three submissions. The same is true of the test account: one properly populated demo tenant serves every reviewer you will ever face.
Run the pre-submission checklist in Anthropic's review criteria before you submit, and be ready for the review to come back with feedback rather than a decision. Track status in the submissions dashboard in your organization settings.
FAQ
No. The submission portal is part of organization settings, which individual plans do not have, so you need a Team or Enterprise organization. Teams routinely hit this at the end of a build, when the person who wrote the server tries to submit from a personal account. Sort out who owns the organization before you write the listing copy.
A missing or incomplete privacy policy, which Anthropic says results in immediate rejection. After that, tools without a title or the applicable readOnlyHint or destructiveHint, since the portal surfaces those before you can submit cleanly. Then test credentials a reviewer cannot actually use, which stalls the review rather than failing it outright.
Anthropic does not publish a service level. The documentation says review times vary with queue volume and the portal is always open. Plan the launch around an unknown wait rather than a date, keep the reviewer contact monitored, and use mcp-review@anthropic.com for escalations if a submission goes quiet.
The slug is permanent once published, so treat it as a product decision rather than a form field. The detail card description you write in the portal is not editable by Anthropic, and the disclaimer shown on connector cards is general rather than customizable per partner.
No. The portal is explicit that inclusion makes the connector discoverable and does not alter the tools it exposes. Anyone can build and run an MCP server for Claude without submitting anything. The directory is a distribution decision, not a technical one.
Desktop extensions are packaged as MCP Bundles and use a separate submission form rather than the portal. MCP Apps, which surface interactive UI, go through the portal but additionally require carousel screenshots: PNG, at least 1,000px wide, three to five images, cropped to the app response with the prompt excluded. Video and GIF are not accepted.
The other surfaces
OpenAI's flow adds a domain-verification challenge and demands test cases up front: at least five positive and three negative. Identity verification gates the whole thing.
Microsoft is a manifest change rather than a portal form. You declare the server in an agentConnectors array and choose between static tool definitions and runtime discovery.
Node8 builds the MCP server, gets the authentication and annotations right the first time, and takes it through review. Most first connectors ship in eight to twelve weeks, and the build is rarely the long pole.