How do you publish an MCP server to the official registry?
Publish to the official MCP Registry (registry.modelcontextprotocol.io) first, because the official docs say it is mainly meant for downstream directory sites. The official docs (read on 2026-11-09) say the registry is a central metadata store for publicly available MCP servers; it stores metadata only, not the packages themselves; it is currently in preview, so breaking changes or data resets may happen; and it does not support private servers. The docs also say it is mainly meant for downstream aggregators, so the name and description you register here can be reused by other directory sites.
1. Publish the package to a public package registry
The registry supports npm, PyPI, NuGet, Cargo, Docker/OCI and MCPB. Remote servers use the remotes field; the address must be publicly reachable, the docs recommend streamable-http, and sse is deprecated.
2. Add the verification string to the package
For an npm package, add mcpName to package.json; for a PyPI package, put a line mcp-name: your-server-name in the README. If you sign in with GitHub, the name must start with io.github.your-username/.
3. Generate and edit server.json with mcp-publisher
Run mcp-publisher init to generate a template, then fill in name, description, version and packages. The name in server.json must match mcpName in package.json.
4. Sign in and publish
Run mcp-publisher login github, then mcp-publisher publish. Afterwards query the registry API once to confirm the entry exists and its status is active.
{
"$schema": "https://static.modelcontextprotocol.io/schemas/2025-12-11/server.schema.json",
"name": "io.github.example/notes-search",
"title": "Notes Search",
"description": "Search and read Markdown notes in a local folder. Read-only; Node.js 18+.",
"repository": { "url": "https://github.com/example/notes-search-mcp", "source": "github" },
"version": "1.2.0",
"packages": [
{ "registryType": "npm", "identifier": "@example/notes-search-mcp", "version": "1.2.0", "transport": { "type": "stdio" } }
]
}
Official moderation is permissive: it only removes illegal content, malware, spam and servers that do not work. The spam examples in the docs include a description stuffed with marketing copy on top of an unrelated implementation, and the same server submitted again under a different name. So write facts in the description, not slogans.
What do directory sites and GitHub say about AI crawlers?
None of the five sites we checked blocks GPTBot, ClaudeBot or PerplexityBot by name. On 2026-11-09 we read their robots.txt with curl, then opened a sample server page on each and looked at the raw HTML: the description on every sample page is written directly into the HTML body, with no script needed.
| Site | What stands out in robots.txt (2026-11-09) | Raw HTML of a sample page |
|---|---|---|
| registry.modelcontextprotocol.io | /robots.txt returns 404, so no rules are declared. The API needs no login: /v0/servers and /v0.1/servers both return 200 and JSON. | The page is a directory entry point and the data comes from the API; we did not sample an entry page. |
| smithery.ai | One group for *: Allow /, with /api/, /_next/, /admin/ and a few others disallowed; no AI crawler is named. | /servers/upstash/context7-mcp: the description is in the body and in the meta description; no JSON-LD. |
| glama.ai | One group for * with Content-Signal: ai-train=no, search=yes, ai-input=yes; sub-pages such as /api/, /mcp/servers/*/tree/ and /blob/ are disallowed; the server page itself is not. | /mcp/servers/shanejonas/openrpc-mcp-server: the README is rendered into the body, and the meta description is the repository's one-line description. |
| mcp.so | One group for *: /admin, /api/, /search and query strings with q= are disallowed; server pages are not. | /servers/sequentialthinking: the description is in the body; it has FAQPage JSON-LD; the meta description is an auto-generated pattern. |
| github.com | GPTBot, ClaudeBot, PerplexityBot and others are named in one group: repository home pages can be fetched, while sub-pages such as /tree/ and /blob/ and ?tab= parameter pages are disallowed. | modelcontextprotocol/servers: the first paragraph of the README is in the body and tells you to go to the MCP Registry for a list of servers. |
This is only what robots.txt says and what our own requests returned, not what an AI crawler really fetches; a CDN or firewall may block on top of that. For what each crawler does, see How do you write robots.txt for AI crawlers?; for whether the body has to be rendered by script, see Do AI crawlers render JavaScript?.
Will an MCP server show up in AI answers?
There are two different things: being connected and used by an AI app, and being recommended or cited in an AI answer. The official docs are clear on the first; for the second we have no evidence.
| Question | What we checked | Why |
|---|---|---|
| 用户怎么把你的服务器接进 Claude | The user enters your remote server URL in the connector settings. The connection is made from Anthropic's cloud, so the server must be reachable from the public internet (Anthropic's IP ranges); a server on an internal network or behind a VPN cannot be reached. | Claude Help Center, “Get started with custom connectors using remote MCP”, read on 2026-11-09 |
| 怎么用在 ChatGPT | Remote MCP servers can be used through plugins in ChatGPT and Codex, ChatGPT deep research and the API; to work with deep research and company knowledge, implement two read-only tools, search and fetch | OpenAI developer docs, “Building MCP servers for plugins and API integrations”, read on 2026-11-09 |
| 目录页会不会被 ChatGPT、Claude、Perplexity 的回答引用 | Not verified: there is no official statement and we have no evidence of actual citations | We did not query any model |
So count the value of a directory page as helping users find and install the server, not as getting AI to recommend you. What an AI answer can cite is still a web page it can read: your README, your own site and the directory page. For how to write them, see {readme}.
How should you write server.json and the first paragraph of the README?
Put the same facts in both places: what the server does, which tools it gives the AI, how to connect and what it requires. The description in server.json is shown as is by the registry and downstream directories, and the first paragraph of the README is rendered as body text by the GitHub page, the npm page and directories such as Glama, so these are the first two passages that people and programs read.
| Weak | Better | |
|---|---|---|
| description | The ultimate AI-powered knowledge platform. Supercharge your workflow! | Search and read Markdown notes in a local folder. Read-only; Node.js 18+. |
| README 第一段 | Welcome to Notes Search, a powerful and flexible solution. | One sentence on what it is, plus the number of tools, transport, runtime and the clients it works with. |
| 工具 | “A range of powerful tools” | List each tool name with one sentence: search_notes searches notes, read_note reads one. |
# Notes Search MCP An MCP server that lets an AI assistant search and read the Markdown notes in a local folder. Read-only, two tools (search_notes, read_note), stdio transport, Node.js 18+. Works with any MCP client. ## Install npx -y @example/notes-search-mcp --dir ~/notes
After that, in order: the install command (one line of npx or uvx), the client config snippet, a table of tools, how authentication works and known limits. For how to write a README overall, see {readme}.
Which directories should you submit to?
Submit to the official registry first, then add third-party directories as needed. Below is what we saw on each site on 2026-11-09; the submission terms are whatever each site's own pages say.
| Directory | What we saw (2026-11-09) | Recommendation |
|---|---|---|
| 官方 MCP Registry | The official docs describe the full publishing flow; the API is publicly readable; moderation is permissive | Must do; do it first |
| GitHub 仓库 | The repository home page and README can be fetched by all three crawlers; the official servers repository says it only holds a small number of reference servers and sends you to the Registry for a list | Write a good README and add topics; do not submit to the official servers repository |
| Glama(glama.ai/mcp/servers) | The page says 118,135 servers, updated 2026-11-09; it has an “Add Server” entry | Search for yourself to see whether you are already listed; if not, submit through its entry; we have not verified its submission terms |
| mcp.so | There is a /submit page; it shows a one-time paid submission option of 39 US dollars, and the page does not make clear how review works without paying | You can submit; whether to pay is your call, and whether paying helps AI citations is not verified |
| Smithery | The page says 214 servers; a banner says the proxy and accounts shut down on 2026-10-31 and that only a public directory supported by Arcade.dev remains; /docs/build/publish returns 404 and we did not find a submission entry | No longer a must-submit directory; if you want to be listed, read its current instructions first |
Each time you submit, note the entry URL and put it in your README and on your own site, so users and crawlers can find it from elsewhere. For being included in “best X tools” articles, see How do I get into “best X tools” list articles?.
Should your own site have a page for each server?
Yes. Directory pages belong to someone else: the description may be rewritten and you do not control the layout or fields. A page on your own site is the version you fully control, and it is a second source an AI can read besides the README.
1. One page per server, with a title that is a question it answers
For example “How do I let Claude search my local Markdown notes?”. Answer in the first paragraph: use Notes Search MCP, two tools, and what environment it needs.
2. Add a table of tools
Tool name, what it does, parameters, read-only or not. This table is the easiest thing to lift from a directory page or README.
3. Add a connection method people can copy
For a stdio server, give the npx command and the client config; for a remote server, give the URL and how authentication works. Write code as text, not screenshots.
4. State the version, date and known limits
Version number, last update date and unsupported cases.
5. Link to the registry entry and GitHub
Link your site page, registry entry and README to each other, and keep the name and description the same.
For how to write documentation pages in general, see How should developer docs be written so AI can understand and cite them?; for package pages, awesome lists and release notes beyond the README, see How do open source projects get recommended by AI search? Six places beyond the README.
FAQ
Should an MCP server be published to the official registry?
Yes. The official docs describe it as the central metadata store for public MCP servers, and downstream directory sites mainly use it. Publishing takes a server.json, a public package or a public remote URL, and the mcp-publisher command line (read on 2026-11-09). It is currently in preview, so the rules may change.
Can AI crawlers read MCP directory sites?
By robots.txt, yes: on 2026-11-09 none of Smithery, Glama, mcp.so and GitHub blocked GPTBot, ClaudeBot or PerplexityBot by name, and the official registry's robots.txt returned 404. Glama writes ai-train=no in its Content-Signal. This is only a judgement from the robots.txt text.
Will directory pages be cited in ChatGPT or Claude answers?
We have no official statement and no evidence of actual citations, so this is not verified. What we can say is that the page description is in the raw HTML, so crawlers can read it.
How should the server.json description be written?
One sentence on what it does, which tools it exposes and the runtime. Official moderation removes servers whose description is stuffed with marketing copy on top of an unrelated implementation, so leave out the slogans.
Can a private or internal MCP server go into the registry?
No. The official docs say the registry does not support private servers, and a server on an internal network cannot be reached by Claude's custom connectors unless you allow Anthropic's IP ranges.
Does submitting to mcp.so cost money?
Its /submit page (as seen on 2026-11-09) shows a one-time paid submission option of 39 US dollars; the page does not make clear how review works without paying, and we have not verified it.
Sources
- Model Context Protocol · The MCP Registry (checked 2026-11-09)
- Model Context Protocol · Quickstart: Publish an MCP Server to the MCP Registry (checked 2026-11-09)
- Model Context Protocol · MCP Registry Supported Package Types (checked 2026-11-09)
- Model Context Protocol · The MCP Registry Moderation Policy (checked 2026-11-09)
- Model Context Protocol · Consuming MCP Registry Aggregators (checked 2026-11-09)
- Anthropic · Get started with custom connectors using remote MCP (checked 2026-11-09)
- OpenAI · Building MCP servers for plugins and API integrations (checked 2026-11-09)
- MCP Registry · GET /v0.1/servers (Read 2026-11-09)
- Glama · glama.ai/robots.txt (Read 2026-11-09)
- Smithery · smithery.ai/robots.txt (Read 2026-11-09)
- mcp.so · mcp.so/robots.txt (Read 2026-11-09)
- GitHub · modelcontextprotocol/servers (Read 2026-11-09)
- kagentx evaluations: the data on the case study pages; each example gives its date.