llms.txt, MCP Server Cards and Agent Standards: What Actually Helps You Get Found by Claude Code in 2026

Which agent standards move the needle for Claude Code? A candid look at llms.txt, MCP Server Cards, AGENTS.md, SKILL.md and real MCP servers, with what the evidence supports and what it doesn't.

Key takeaways

  • Claude Code does not read llms.txt as instructions. It reads CLAUDE.md (or AGENTS.md if there is none). llms.txt only helps when Claude Code fetches your docs through live retrieval, which usually happens after a person points it at your site.
  • MCP Server Cards reached Final status as an MCP Extensions Track proposal (SEP-2127). I found no evidence that Claude Code uses them for discovery today, so treat them as cheap future-proofing, not a growth lever.
  • A working MCP server is the strongest agent-native asset you can ship, because it makes you callable instead of merely quotable.
  • Clean markdown, accessible crawlers and accurate docs matter more than any single file. Check that Claude-User and ClaudeBot aren't blocked.
  • Measure by asking Claude Code your customers' questions and recording the answers before and after each change.

The honest starting point

There is a lot of confident advice about "agent standards" right now, and much of it is untested. Evil Martians published one of the more candid takes in November 2026. Their team found that llms.txt, which they had ranked as their top pick in April, is still read by almost no AI tool. In their traffic, named AI assistants made 47 requests for the file, and 43 of those came from Claude Code after a person pointed it at their site.

That last detail matters. It means llms.txt does get used by Claude Code, but mostly in a human-initiated, on-demand way. Nobody is discovering you through it yet.

Evil Martians article on which techniques make a site visible to LLMs and AI agents

So the useful question is not "which standards should I adopt?" It is "which parts of Claude Code's decision process can I actually influence, and which standard touches each one?"

How Claude Code decides what to use or recommend

A May 2026 analysis from Flywheel describes Claude Code as blending four things: training-data priors, live documentation retrieval, callable MCP servers, and project-level instructions. In practice the flow looks like this:

  1. Recall what the model already knows from training.
  2. Retrieve current information with WebSearch and WebFetch when needed.
  3. Call MCP servers that expose tools and documentation.
  4. Follow project instructions such as CLAUDE.md.
  5. Synthesize all of that and recommend or use a tool.

Each standard maps onto one of those steps. That is a better way to prioritize than a checklist.

StandardWhich step it touchesEvidence Claude Code uses itEffort
llms.txtRetrieve (WebFetch)Used on demand, mostly after a person points Claude Code at your siteLow
llms-full.txtRetrieveSame as above, one big file instead of an indexLow
Markdown versions of pagesRetrieveClaude Code sends Accept: text/markdown and takes what you serveLow to medium
MCP serverCallDirect, documented supportMedium to high
MCP Server CardDiscovery before connectionNone found for Claude Code todayLow
CLAUDE.md / AGENTS.mdProject instructionsRead from the user's own repo, not your siteLow (for your own repo)
SKILL.mdReusable workflowsLoaded when users install the skillMedium

Notice that three of these live in someone else's repository or configuration. You influence them only indirectly, by shipping something worth installing.

llms.txt: cheap, harmless, overrated

llms.txt is a Markdown file at your site root that gives AI systems a curated map of your most important content. Jeremy Howard's rationale, quoted by Evil Martians, was that we need something like robots.txt but designed to help AI systems understand and use website content. The llms.txt proposal is explicit that it is for inference, not training: an agent reads it on demand while helping a user, then follows the links.

The v2 proposal also suggests serving a clean markdown version of each page at the same URL with .md appended, and using rel="alternate" type="text/markdown" and rel="describedby" link relations so clients can find those files.

Does Claude Code read it?

Not as instructions. Moxie Docs puts it plainly: Claude Code reads CLAUDE.md, or AGENTS.md if there is none. llms.txt enters the picture only when the agent fetches your documentation. The Flywheel analysis says the same thing from the other side: a clean llms.txt and a well-structured README improve the quality of what live retrieval returns, which can improve both accuracy and the likelihood of a recommendation. It is not a ranking signal in the search-engine sense.

My read: build it, because it takes an afternoon and the same file helps several agents. Do not build your roadmap around it.

What a useful one looks like

# Acme API

> Payments API for marketplaces. REST, webhooks, SDKs for Node and Python.

## Docs
- [Quickstart](https://acme.dev/docs/quickstart.md): first payment in 10 minutes
- [Authentication](https://acme.dev/docs/auth.md): API keys and OAuth
- [Webhooks](https://acme.dev/docs/webhooks.md): events and retries

## Optional
- [Changelog](https://acme.dev/changelog.md)

Link to markdown versions, not marketing pages. Keep descriptions factual. Skip the sales copy, since a coding agent is trying to get an API call right, not be persuaded.

If you want to generate one quickly, Promptwatch has a free llms.txt generator.

MCP Server Cards: real standard, unproven payoff

This is the one most likely to be oversold, so here is exactly what is established.

SEP-2127, "MCP Server Cards - HTTP Server Discovery", has Final status as an Extensions Track SEP, created January 21, 2026. The SEP page is the source of truth, and it points to the experimental-ext-server-card repository for the normative schema, discovery mechanics and security analysis.

What the SEP says:

  • A Server Card is static, pre-connection metadata describing a remote MCP server.
  • Cards can be hosted at any unreserved URI, with <streamable-http-url>/server-card as the recommended location.
  • Domain-level discovery goes through an AI Catalog, for example /.well-known/ai-catalog.json, which links to or embeds cards.
  • Cards cover remote servers only. The MCP Registry's server.json covers registry entries, including locally installable packages.
  • If a card and the live server/discover response disagree, clients should prefer the live values.
  • Stated use cases: IDE autoconfiguration, crawlers indexing MCP servers across domains and registries, and showing metadata without opening each endpoint.

Two cautions. First, the location changed from earlier proposals. SEP-1649 suggested /.well-known/mcp/server-card.json, and SEP-1960 suggested a /.well-known/mcp manifest. Tutorials written around early 2026 that use those paths may be stale. Check the repo before copying anything.

Second, and more important: I found no evidence that Claude Code reads Server Cards for discovery today. The stated use cases are about tooling that indexes and configures servers, not about an agent deciding what to recommend mid-conversation. Publishing a card is low effort and plausibly useful as the ecosystem matures. Just don't expect a bump in recommendations from it this quarter.

MCP servers: the one that changes behavior

MCP is an open standard that lets an agent use outside services. Fragments Studio uses a kitchen analogy: the model is the chef, the harness is the kitchen, and MCP servers are outside suppliers that do work elsewhere and return the result. Anthropic introduced the protocol and Claude supports connecting MCP servers directly.

This is where the distinction between being quotable and being callable becomes real. Text-oriented GEO tactics make your page easier to cite. An MCP server lets Claude Code use your product during the task, which is a much stronger position. Flywheel calls Claude Code's MCP-plus-Context7 pairing its distinctive agent-native packaging path.

Practical priorities if you ship one:

  • Expose tools that map to real jobs your customers do, not a 1:1 mirror of your REST API.
  • Include documentation access so the agent can answer questions about your product without leaving the session.
  • Track call volume, and watch for developers arriving already naming your tool.
  • List it somewhere discoverable: the MCP Registry, your docs, and your llms.txt.

A small example of the last point: there is even a community MCP server that lets agents fetch and search llms.txt files, so the two standards connect more than you might expect.

AGENTS.md, CLAUDE.md and SKILL.md

The Hugging Face overview from October 2026 groups these as instruction formats: AGENTS.md for repository instructions, SKILL.md for reusable workflows, and CLAUDE.md for Claude Code project instructions. Support varies across agents.

For your own product, the lever is adoption by users. If developers put a line about your SDK in their CLAUDE.md, Claude Code will follow it. You can encourage that by publishing a short, copyable snippet in your docs. A published skill is similar: Evil Martians wrote about publishing agent skills like packages, with a discovery index, digests and install sources.

I haven't verified how every agent loads these files, so check each vendor's current docs rather than trusting a summary, including this one.

Make sure Claude can reach you at all

Standards are pointless if the crawler is blocked. Promptwatch Data tracked Claude-User, the crawler used when someone asks Claude a question or uses Claude Code, growing from about 30 visits a day in mid-December 2025 to several thousand a day by mid-April 2026. That is more than 100x. Its peak daily share of tracked AI citation crawler traffic was 1.73% on April 12, and its share over the 130-day window was 0.45%.

So Claude is a small but fast-growing channel, one Promptwatch describes as worth unblocking and monitoring rather than pouring budget into. For context, Promptwatch's AI crawler traffic data shows OpenAI at 79.8% of verified AI crawler requests in the week of August 31 to September 6, 2026, down from 94.8% in early June.

The checklist is short:

  1. Confirm Claude-User and ClaudeBot aren't blocked in robots.txt or by CDN and WAF bot rules.
  2. Baseline your Anthropic user-agent hits in server logs and review monthly.
  3. Serve markdown when a client sends Accept: text/markdown. Claude Code does, and takes whatever you negotiate.
  4. Make sure your docs render without JavaScript, since WebFetch is not a browser.

Tools like Promptwatch can show IP-verified crawler hits per page, which helps separate an access problem from a content problem. Its AI crawler logs cover Claude-User, ClaudeBot and others, and its Agentic Readiness Check tests which agent standards your site supports.

Favicon of Promptwatch

Promptwatch

AI search visibility and optimization platform
View more
Screenshot of Promptwatch website

Measuring whether any of this works

Agent tool selection is young and badly instrumented, so be skeptical of anyone selling a dashboard number. Two methods hold up:

  1. Direct testing. Ask Claude Code the questions your customers would ask, record whether and how it recommends you, and re-run after each change such as shipping an MCP server or fixing a blocked crawler. Change one thing at a time.
  2. Proxies. Track MCP call volume, installs and sign-ups from developers who arrive knowing your tool by name, and check how text answer engines describe you.

Promptwatch also tracks how coding agents like Claude Code, Codex and OpenCode recommend tools and SDKs, which is useful if you sell to developers. For other options, the GEO software directory lists alternatives.

What to do, in order

PriorityActionWhy
1Unblock Claude-User and ClaudeBot, review logsNothing else works without access
2Fix docs: clean markdown, no JS wall, accurate and currentFeeds live retrieval
3Publish llms.txt linking to markdown pagesCheap, helps several agents
4Ship an MCP server with docs and real toolsMakes you callable
5Publish a Server Card per the current repo specFuture discovery, unproven today
6Provide CLAUDE.md snippet and a skillLets users wire you in
7Test with real prompts after each changeThe only direct signal

The bottom line

The boring work wins. Accessible, accurate, markdown-friendly docs and a genuinely useful MCP server affect what Claude Code does. llms.txt is a reasonable supporting file and Server Cards are a bet on where discovery goes. Neither is a shortcut. If someone promises that a single file will get you recommended by Claude Code, ask for the logs.

If you'd rather hand the SEO and GEO side of this to specialists, 1001 SEO Media runs AI search programs that cover technical access, content and measurement.

Share:

© 2026 Toolsolved · Find the best marketig tools · RSS

Toolsolved is an affiliate review site. When you click links to vendors or buy through links on our site, we may earn an affiliate commission at no extra cost to you.

Toolsolved is a review website based on user reviews on Reddit and G2, and on publicly available information. We keep everything as up to date as possible, but pricing and features can change. Always confirm the details with the vendor before purchasing.