How to Get Your AI Agent or MCP Server Discovered
Updated 2026-11-18
To get an AI agent or MCP server discovered, publish it to the official MCP Registry, keep a stable remote endpoint, write clear tool names and descriptions, and add documentation that answers common questions in plain sentences. Then add optional signals such as llms.txt and an A2A agent card, and list the server in directories. Keep it online, because dead endpoints get noticed.
1. Publish to the official MCP Registry
The official registry is the main source that other directories read. If you do one thing, do this. It is free, and a listing there is picked up by downstream aggregators on their own schedule. Agents Index, for example, imports from it daily. The steps are in how to publish an MCP server to the official registry.
Use a namespace that matches your identity. A reverse-DNS name verified by DNS or HTTP shows that the publisher controls the domain. A GitHub name (io.github.you/...) shows control of a GitHub account.
2. Run a stable remote endpoint
Web clients such as ChatGPT and claude.ai can only connect to remote servers. A local npm package alone leaves those users out. Offer a Streamable HTTP endpoint at a path that does not change, such as https://mcp.yourcompany.com/mcp, and keep the URL permanent.
Think about sign-in. Agents Index probe data, Nov 2026, shows how the registry's remote endpoints split: of those probed, 23,704 answered without sign-in, 13,608 were online but required sign-in (OAuth or API key), and 7,189 were unreachable. If your server needs OAuth, support the standard flow so clients can discover it. See the State of MCP report for the full numbers.
3. Write good tool names and descriptions
The model reads your tool list to decide what to call. Weak names and descriptions mean the tool is skipped or misused. The MCP documentation's own example prefers a name like calculator_arithmetic over just calculate.
- Use specific names. The spec suggests letters, digits, underscore, hyphen and dot, with no spaces.
- Say what the tool does and when to use it, in one or two sentences.
- Describe every parameter, with example values.
- Set
readOnlyHinton tools that only read data, so clients can avoid unnecessary approval prompts. - Keep the tool count small. A server with 80 vague tools is harder to use than one with 10 clear ones.
{
"name": "list_invoices",
"description": "List invoices for the signed-in account. Use this to answer questions about billing history. Read-only.",
"inputSchema": {
"type": "object",
"properties": {
"status": { "type": "string", "enum": ["open", "paid", "void"], "description": "Filter by invoice status" },
"limit": { "type": "integer", "description": "Maximum results, for example 20" }
}
}
}
4. Add llms.txt
llms.txt is a proposal for a markdown file at /llms.txt that gives AI systems a curated overview of your site. The format is an H1 with the project name, a blockquote summary, and then H2 sections with lists of links:
# Acme Billing
> Acme Billing is an invoicing API and a remote MCP server for reading invoices and customers.
## Docs
- [MCP server setup](https://acme.example/docs/mcp): endpoint URL, auth and client config
- [Tool reference](https://acme.example/docs/mcp-tools): every tool with parameters
## Optional
- [Changelog](https://acme.example/changelog)
It is a proposal, not a standard that every engine follows, and we cannot confirm that any particular answer engine reads it. It takes ten minutes, and it also gives human readers and agents one index of your key pages.
5. Publish an A2A agent card
If your product is an agent that other agents can talk to, publish an Agent Card for the A2A protocol. The A2A discovery docs say the standard location is:
https://your-domain.example/.well-known/agent-card.json
Clients send a GET request to that path. The card describes your agent: identity (name, description, provider), the service endpoint (url), capabilities such as streaming, required authentication, and a list of skills with an id, name and description. Check the current schema at a2a-protocol.org before you publish, because field names can change between spec versions. Do not put secrets in the card. Our MCP vs A2A guide explains when you need one.
6. Write docs that can be quoted
Answer engines quote short, direct passages. Structure your docs the same way:
- Open each page with a two or three sentence answer to the question in the title.
- Use headings that match real questions: "How do I connect X to Claude Code?"
- Put exact commands and config in code blocks, one per client.
- State plain facts: the endpoint URL, auth method, rate limits, pricing model.
- Put a date on the page and keep it current.
Write one page per client you support. Searchers look for "your product plus Cursor" or "your product plus ChatGPT", and a page for each matches that.
7. List in directories
The official registry is the base. Beyond it, list in directories where your users look. Options include client marketplaces (the Cursor Marketplace, for one), community lists, and independent directories like Agents Index. We list servers from the official registry for free, and you can also submit a listing directly. Free listings are included, and there is an optional paid fast-track review if you want a quicker look. Advertising is another option, but it is not required, and none of the other steps depend on it.
8. Keep it up
A listing that points to a dead endpoint costs you trust. Directories that probe servers, as we do, show whether an endpoint answers. Monitor your own URL, return the correct auth challenge (a 401 with discovery metadata) instead of a 500 error, and avoid blocking probe traffic with bot or geo rules. See how we probe endpoints.
Checklist
- Server published to the official registry with a verified namespace.
- Stable remote endpoint with documented auth.
- Tool names, descriptions and parameters reviewed by someone who did not write them.
/llms.txtlive.- Agent card at
/.well-known/agent-card.json, if you run an A2A agent. - One docs page per supported client, each opening with a direct answer.
- Listed in the directories your users use.
- Uptime monitoring in place.
Sources
- The MCP Registry - modelcontextprotocol.io
- Registry quickstart - modelcontextprotocol.io
- MCP specification: Tools - modelcontextprotocol.io
- MCP architecture overview - modelcontextprotocol.io
- The /llms.txt file - llmstxt.org
- Agent Discovery - a2a-protocol.org
FAQ
What is the first thing to do to get an MCP server discovered?
Publish it to the official MCP Registry. Many directories, including Agents Index, pull from it, so one listing there is picked up automatically by downstream sites.
What is an A2A agent card?
An Agent Card is a JSON document that describes an agent: its name, endpoint, capabilities, authentication requirements and skills. A2A clients can fetch it from /.well-known/agent-card.json on your domain.
Does llms.txt help AI answer engines find my agent?
llms.txt is a proposal for a markdown file at /llms.txt that gives AI systems a curated overview of a site. No source we reviewed confirms that any specific answer engine reads it, but it costs little to add.
Does uptime affect discovery?
Yes. Directories that probe endpoints, as Agents Index does, show whether a listing responds. In Agents Index probe data from Nov 2026, 7,189 remote endpoints in the registry were unreachable.