Submit your MCP server to AI agent catalogs with Distribute

Distribute gets one MCP server listed in AI agent catalogs. Every catalog asks for the same facts under different names and limits. In Distribute you answer them once, and MCP Tower reuses the answers for every catalog. It then checks your server against each catalog’s rules, prepares what each catalog takes, tracks each submission’s status, and emails you about approvals, changes and drift.

Open Distribute in the main menu and pick a server.

The catalogs

Catalog	What MCP Tower prepares	Status
Official MCP Registry	`server.json`, validated and published by MCP Tower on your click	Automatic
ChatGPT (OpenAI plugins)	The plugin bundle (ZIP), domain verification and the next steps	You set it
Claude (Connectors Directory)	The answer to every portal question, in portal order	You set it
Muse (Meta)	The answers to both Muse forms	You set it
Hermes	The catalog entry, sent as a pull request from your GitHub account	Automatic
Grok (xAI form)	The 12 form answers	You set it
Grok Build and Cursor	Two pull requests from your GitHub account, one per repository	Automatic
Perplexity	The form’s answers and a link to the form with the answers filled in	You set it
Google Gemini	Coming soon	–
Microsoft Copilot	Coming soon	–

The Gemini and Copilot pages show what those catalogs will need and which of it your answers already cover.

Overview and catalog pages

The left menu starts with Overview, which shows how many answers you have given, for example “38 of 46 answers”. Below it, each catalog has a mark:

missing: not submitted yet, or not approved. A sub-label says why, for example “5 answers missing”, “2 checks failing”, “Ready to submit” or “Not approved”.
reviewing: submitted and waiting for the catalog, including when it asked for changes.
published: listed. A warning dot means your server changed since the submission and the catalog needs it again (see Status, drift and emails).

Overview has a readiness card per catalog with its next action, then every check, then the questionnaire. Each part of the questionnaire can be edited in place.

Each catalog page has five sections: what’s missing, the catalog’s checks, its output (answers, a file, a bundle, a pull request or a publish), the next steps, and the status with its history. Each page also says when MCP Tower last checked the catalog’s rules, with a link to the source.

The wizard

The first time an Admin or Editor opens Distribute for a server with no saved answers, a wizard walks through the questionnaire:

Where to list. Every catalog is picked. Fields that none of your picked catalogs use are hidden.
Your organization. Skipped when you already filled it in for another server.
Listing. Name, subtitle, summary, description, categories, capabilities and example prompts.
Branding. Icon, color, screenshots and demo video.
Links. Website, documentation, support, privacy policy and terms. Each link is checked when you leave the field.
Connection and access. The URL, authentication and tools, from your latest scan, and the test accounts.
Review cases. Five requests your server should handle and three it should decline.
Data and privacy. Where data is stored, what you collect, retention and deletion.
Summary. The readiness of each catalog and what to do first.

Each step saves as you go. You can leave and come back, and a teammate can continue where you stopped. The Wizard button next to the page heading starts again at step 1 and keeps every answer.

What MCP Tower already knows is filled in for you: the URL, authentication, tools and icon from your latest scan, your organization’s answers, and links and data answers from your other servers. A filled-in value says where it comes from. Every field explains what it is and why catalogs ask for it, names the catalogs that use it, and counts characters against the strictest limit among your picked catalogs.

You can override a text answer for one catalog, for example a shorter description for Grok, without changing it for the others.

Checks

MCP Tower checks your server, its tools, your answers and your assets against each catalog’s rules. The checks run after every scan, including Inspect’s daily rescan, and after you save answers. Each check has a level for each catalog:

Blocker: the catalog rejects a server that fails it.
Warning: the catalog’s reviewers are likely to ask about it.
Tip: advice that helps agents use your server.

Overview shows every check, failing ones first. A catalog page shows only that catalog’s checks. A check about tools names the tools that fail it.

Run checks evaluates every check again on your latest scan, and fetches your links, OAuth metadata and icon again. It doesn’t scan your server: to rescan, use Scan again in Inspect.

Some checks are heuristics and can be wrong about your server. When a failing check, or one tool in it, doesn’t apply, press Dismiss and say why. The check then shows as dismissed, with who dismissed it.

How each catalog is submitted

Official MCP Registry

MCP Tower generates server.json, validates it with the registry and shows any problem next to the answer that causes it. Publish to registry publishes it, only when you press it. Each publish gets a new version.

The registry only publishes under a name you prove you own. By default MCP Tower proves your server’s public hostname for you (for example mcp.customer.com): MCP Tower’s middleware serves the proof on that hostname, so you change nothing. The name is then your hostname reversed, for example com.customer.mcp/orders. You can instead prove your whole domain with a DNS TXT record, for a shorter name such as com.customer/orders. The page shows the record to add at your DNS provider and Check DNS looks it up. Choose the method before the first publish: switching later lists the server under a new name.

The page also prints the equivalent mcp-publisher commands, for teams that publish from CI.

ChatGPT

Generate bundle builds the ZIP that OpenAI’s Plugins page takes. Regenerate builds a new version from your latest scan and answers, shows what changed since the last submitted bundle, and asks for release notes (at most 500 characters). Every bundle is kept and can be downloaded again.

The next steps walk you through OpenAI’s upload. For domain verification, paste OpenAI’s token into MCP Tower: its middleware serves the token on your server’s public hostname, so nothing changes on your origin server. The test account is printed in the review step, because it doesn’t go in the ZIP.

Claude, Muse, Grok and Perplexity

These catalogs take a form. MCP Tower prints every answer in the form’s order, with its limit and a copy button, and links to the form.

Claude: the ten steps of the Connectors Directory portal, the icon to download, and the screenshots when your server has widgets. The seven compliance statements are listed for you to read: you make them yourself in the portal.
Muse: the “Credentials” and “Data processing” pages, with Muse’s own option labels.
Grok: the 12 answers of xAI’s form. That form can also list you in Cursor and Grok Bot. If you use the Grok Build and Cursor pull requests, pick only Grok in the form, so Cursor doesn’t get the same server twice.
Perplexity: every answer by section, and a link that opens Perplexity’s Google Form with the short answers filled in. Sign in to Google, add the answers the link leaves out, and submit. The brand asset page the form asks for is a public, unlisted page with your server’s icons.

You submit each form yourself. Then set the status on the catalog page: Submitted, Changes requested, Not approved or Published, which asks for the listing URL.

Hermes, Grok Build and Cursor

These catalogs take pull requests to their GitHub repositories, opened from your own GitHub account:

Hermes: a manifest in NousResearch/hermes-agent.
Grok Build: a plugin in xai-org/plugin-marketplace.
Cursor: a plugin in cursor/plugins. Pick Grok Bot too to be listed there from the same entry.

Press Sign in with GitHub once, then Open pull request. MCP Tower forks the repository into your account, commits the files on a branch, and shows you the diff, title and description. The pull request opens only when you confirm. Later changes update the same entry.

Your GitHub connection is listed under Connected accounts in your account menu, where you can disconnect it. MCP Tower uses it only for pull requests you confirm.

You can also open these pull requests with your AI agent: see Use MCP Tower from your AI agent.

Status, drift and emails

MCP Tower follows the registry and the pull requests itself:

Registry: published while the entry is active. Deprecated or deleted turns it back to missing, with an email.
Pull requests: open is reviewing, merged is published, closed without merging is missing, with an email. A pull request opened elsewhere, for example by your agent, is found the same way.

The other catalogs change status when you set it. Every change is kept in the page’s history, with who made it and when.

Each submission keeps a snapshot of your tools and answers. When a later scan finds new, removed or changed tools, or your answers changed, the catalog page shows what changed and suggests resubmitting: publish a new registry version, open an update pull request, or regenerate the ChatGPT bundle.

Admins and Editors get an email when a status changes (unless they changed it), when a server is published, and once when drift appears.

Draft answers with AI

Some answers take long to write. The AI buttons ask Claude, Anthropic’s model, to draft them from what MCP Tower knows about your server:

Button	Drafts
Draft with AI (Review cases)	Five positive and three negative review cases
Draft with AI (Listing)	Subtitle, summary, description, capabilities, example prompts and more
Generate from privacy policy	Data retention and deletion process
Draft release notes	ChatGPT release notes, from what changed since the last submitted bundle

A draft only runs when you press a button, and nothing is saved until you press Accept. You can also edit the draft or discard it. An accepted draft is marked “Drafted with AI” until someone edits it. Drafts follow the same limits and checks as typed answers.

Answers drafted from your privacy policy quote the sentences they are based on, so you can check them. When the policy doesn’t say, the answer is “Not stated in your privacy policy” and the field stays empty.

Each organization gets 100 drafts a day. A draft that fails validation twice doesn’t count.

What is sent to Anthropic. Your server’s instructions, tools, resources and prompts from the latest scan, your answers, the text of your website’s home page and privacy policy, and usage counts: the most-called tools and the agents that call them over the last 30 days. MCP Tower never sends test-account passwords, secrets, event payloads or recorded argument values.

Test accounts

Most catalogs need a test account when your server requires sign-in. Passwords are stored encrypted. Reviewers, and the AI agents you use to submit, see the password, so use a dedicated account with limited access and sample data only, never a real user or an admin, and without MFA. Change its password after review. Every time someone reveals a password, it is recorded in your audit log.

Viewers and Editors

Your role	In Distribute you can
Viewer	See every page, answer, check, status and generated file; copy answers; download bundles and `server.json`; see test-account usernames
Editor or Admin	Also use the wizard, edit answers, add test accounts and reveal passwords, use AI drafting, run and dismiss checks, generate bundles, publish to the registry, open pull requests and set statuses

A Viewer who opens a server with no saved answers sees the known values and the note that an Admin or Editor can fill in the rest.