publish-mcp-server
Skill1 starsChanged 9 days agoMITView on GitHub
Start here to take one MCP server into both the ChatGPT plugin directory (platform.openai.com/plugins) and Claude's connector directory (claude.ai/directory) end to end. Use whenever someone wants to publish, list, submit or "ship" an MCP server, ChatGPT app, ChatGPT plugin, Claude connector or MCP Apps UI, asks what is left before submitting, or is stuck at any step from first setup to a reviewer's findings, even if they only name one of the two stores.
In this repository8 other files it gives its agents.
- AGENTS.mdAGENTS.md
- mcp-app-cardsSkill
- mcp-server-for-pluginsSkill
- plugin-oauthSkill
- plugin-packageSkill
- plugin-prerequisitesSkill
- submit-to-chatgptSkill
- submit-to-claudeSkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 32
Find the stage: Ask only what you cannot see.
- Distinctive: line 62
Read the failure literally before theorising.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Publish one MCP server to ChatGPT and Claude
Both stores now run on the same thing: a remote MCP server over Streamable HTTP, OAuth sign-in, and (optionally) MCP Apps cards. So one server can be listed in both. What differs is a handful of rules that conflict, two packaging formats and two review portals. This skill finds where the person is and runs the right next skill, in this order:
| Stage | Skill | Done when |
| 1 | plugin-prerequisites | accounts, domain, legal pages, support mailbox, reviewer account exist |
| 2 | mcp-server-for-plugins | every tool has a verb_first name, a title, all four hints and an output schema; results fit the size budget |
| 3 | plugin-oauth | discovery documents, PKCE, CIMD and DCR work; a reviewer can sign in with no codes |
| 4 | mcp-app-cards (skip if no cards) | the stand-in host test passes in all three modes; a card renders in both hosts |
| 5 | plugin-package | one listing source writes both packages; the package checker passes |
| 6 | submit-to-chatgpt | the portal shows no blocking findings; review submitted |
| 7 | submit-to-claude | the connector portal's ten steps are filled; review submitted |
Stages 6 and 7 are independent once 1–5 are done; most people submit to Claude first because it needs fewer assets (no demo video).
How to run it
- Find the stage. Ask only what you cannot see. If you are in the server's repo, look: is
there an
/mcproute,/.well-known/oauth-authorization-server,ui://resources, aplugin.jsonor.claude-plugin/? Each answers a stage without a question. - Run the stage's skill, then come back here and tick it off.
- Keep one listing source. Names, descriptions, prompts and test cases live in one file and both packages are generated from it (stage 5). Writing two by hand is how listings drift.
- Ask before anything leaves the machine: a push that deploys, a portal submission, a public repo. Each store review takes days to weeks, so a rejection costs far more than a check.
The conflicts to know up front
These are the places where doing the right thing for one store breaks the other. Each skill covers its own in detail; know they exist before starting.
- Card domain. Claude shows a card only if
_meta.ui.domainis its own hash of your URL or is absent; ChatGPT wants a domain declared. Leaveui.domainout and give ChatGPT its alias,openai/widgetDomain(mcp-app-cards). - Card start-up. ChatGPT also injects
window.openai. A card that skips the MCP Apps handshake when it seeswindow.openaiworks in Claude and times out in ChatGPT after about 5 seconds (mcp-app-cards). - Tool choice. ChatGPT has its own features (quizzes, web search). Without server
instructions that say when to use your tool, it may answer a test prompt without calling you
(
mcp-server-for-plugins). - Sign-in. Both register as OAuth clients by Client ID Metadata Document (CIMD) from their
own domains; other apps use Dynamic Client Registration (DCR). Support both (
plugin-oauth). - Listing words. ChatGPT allows 30 characters for a short description; Claude's card shows
about 90 characters of a one-liner. Write per store from one source (
plugin-package).
When something fails
Read the failure literally before theorising. A portal finding, a card that says "timed out", a sign-in that bounces: each skill lists the messages it has seen and what fixed them. If the message is new, reproduce it against a stand-in (the cards skill ships a stand-in host) before changing code that already works in the other store.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

