# llms.txt vs MCP

> llms.txt is a static Markdown briefing an agent reads; MCP is a live protocol an agent calls. How the two differ and when a site or product needs each.

Source: https://metavert.io/compare/llms-txt-vs-mcp  
Published: 2026-10-07  
Updated: 2026-10-07

Comparison

[llms.txt](https://metavert.io/llms-txt) and the [Model Context Protocol (MCP)](https://metavert.io/model-context-protocol) are the two most common answers to the same question: how does a website or product make itself usable by AI agents? They are frequently presented as rivals. They are better understood as different layers. llms.txt is a static Markdown file that tells an agent what a site contains and where to read more. MCP is a protocol through which an agent calls a live server to fetch data or take actions.

The short version: llms.txt gives an agent *context*, and MCP gives it *tools*. Those are two of the four pillars of [agent experience](https://metavert.io/agent-experience), and most products that take agents seriously end up shipping both. The useful questions are which to build first, what each costs to maintain, and what the evidence says either one achieves.

## Feature Comparison

| Dimension | llms.txt | MCP |
| --- | --- | --- |
| What it is | A Markdown file at `/llms.txt` summarizing a site and linking to key pages | An open protocol connecting AI applications to external data, tools and workflows |
| Origin and governance | Proposed by Jeremy Howard, September 2024; an informal community proposal | Created by Anthropic; governed since December 2025 by the Linux Foundation's Agentic AI Foundation |
| Nature | Static content, fetched over ordinary HTTP | Live client-server interaction with typed tool calls |
| What an agent can do | Read | Read, query and act (create, update, provision, purchase) |
| Agent-side setup | None; any agent that can fetch a URL can use it | The server must be connected in an MCP-capable client, usually by the user |
| Authentication | None; public by design | Supported, so a server can expose private or per-user data |
| Freshness | As current as the last regeneration | Live at call time |
| Cost to build | Low; often generated automatically by a docs or CMS platform | Moderate to high; a service to design, host, secure and version |
| Discovery | Well-known path at the site root | The [MCP Registry](https://metavert.io/mcp-registry) and downstream marketplaces |
| Main risk | Staleness; a wrong map misdirects every reader | Security; tool access and untrusted tool output |
| Evidence of use | 97% of files got zero requests; coding agents are the main readers (Ahrefs, June 2026) | "More than 10,000 published MCP servers" (Linux Foundation, December 2025); trackers now report far more |

## Detailed Analysis

### A Document Versus an Interface

The llms.txt proposal describes itself as a way "to provide information to help agents use a website." The file has a fixed, simple shape: a title, a one-paragraph summary, and sections of links, ideally pointing to clean [Markdown](https://metavert.io/markdown) versions of each page. An agent reads it the way a person reads a table of contents, then fetches what it needs. Nothing about the exchange is interactive, and the site learns nothing about what the agent wanted.

MCP is defined by its project as "an open-source standard for connecting AI applications to external systems." A server advertises tools with names, descriptions and parameters; the agent decides which to call and receives structured results. That makes MCP capable of things a file cannot do: answer a query against live data, return only the relevant fragment, respect a user's permissions, or change state. It also makes MCP an application to operate, not a file to publish.

### Who Actually Reads Them

For llms.txt the evidence is unusually direct. Ahrefs analysed logs for 137,210 domains in May 2026 and found that 97% of llms.txt files received no requests at all. Among the AI traffic that did arrive, coding agents and agent infrastructure were the largest category at 10.5% of requests, while AI retrieval bots accounted for 1.1%, and Ahrefs observed that no AI bots go looking for the file where it does not exist. The implication is that llms.txt is not, on current evidence, a lever for visibility in [AI search](https://metavert.io/ai-search). Its demonstrated audience is [coding agents](https://metavert.io/ai-coding-agents) working against documentation.

For that audience it does measurable work. Mintlify, which hosts developer documentation, reported in July 2026 that serving llms.txt reduced agent error rates by almost 90% in its benchmark, since agents stopped guessing at URLs. The figure is vendor-reported and specific to docs sites, but it matches the mechanism: a map prevents wrong turns.

MCP usage is harder to measure from outside because calls happen inside authenticated sessions. Supply is clearly large. The Linux Foundation cited more than 10,000 published servers in December 2025, and third-party directories checked in October 2026 listed between roughly 22,000 and 97,000 depending on how they count. Supply is not demand, though, and there is no public equivalent of the Ahrefs log study showing how often typical MCP servers are called.

### Cost, Maintenance and Risk

llms.txt is cheap to publish and cheap to get wrong. Generated from the content system, it stays current on its own; written by hand, it drifts, and a stale file confidently points agents at pages that have moved. The security surface is negligible because it exposes nothing that was not already public.

An MCP server is a product surface with the obligations that implies: authentication, rate limits, versioning, and a decision about which actions an agent should be allowed to take without confirmation. Because models treat tool descriptions and results as input, MCP also inherits [prompt injection](https://metavert.io/prompt-injection) risk in both directions. A read-only server narrows that exposure considerably and is a sensible first step for content sites.

### Reach and Discovery

llms.txt needs no cooperation from the agent's operator. Any agent with a fetch tool can read it during a task, which is why it works for anonymous, first-contact use. MCP generally requires someone to have connected the server beforehand, so it serves existing users and deliberate integrations better than cold discovery. Registries and marketplaces are narrowing that gap, and agents can also learn that an MCP server exists from the documentation they read, which is one reason the two approaches reinforce each other: the llms.txt file is a natural place to state that an MCP endpoint exists and how to connect to it.

### In Practice

The common sequence is content first, tools second. Publish Markdown copies of pages and an llms.txt generated from them, confirm in server logs that agents are fetching them, then add an MCP server where agents need to query or act instead of merely read. The division between the two is increasingly handled by the publishing platform: [LightCMS](https://metavert.io/lightcms), the CMS serving this site, generates llms.txt and Markdown copies of every page and also exposes a public read-only MCP endpoint for visitors' agents, so the same content is available through both routes.

## Best For

#### Documentation for a developer tool

Both

llms.txt and Markdown pages cover the coding agents that fetch docs mid-task; an MCP server adds search across docs and account-aware actions for connected users.

#### Marketing or editorial site

llms.txt

The content is public and read-only. A generated llms.txt is nearly free, though the Ahrefs data suggests modest expectations for traffic.

#### Letting agents take actions in a product

MCP

Creating records, provisioning resources or changing settings requires authenticated tool calls. A static file cannot do any of it.

#### Private or per-user data

MCP

llms.txt is public by definition. MCP supports authorization, so results can respect each user's permissions.

#### Smallest possible first step

llms.txt

One generated file and Markdown page copies, with no service to run, no security review and nothing for the agent's user to install.

#### Large or fast-changing catalogs

MCP

Prices, inventory and status change faster than a file regenerates, and a query returns the relevant rows instead of the whole catalog.

#### Improving AI search citations

Neither

There is no evidence that either one raises citation rates in AI search. Ahrefs found AI retrieval bots made up 1.1% of llms.txt requests.

#### Reaching agents on first contact

llms.txt

Any agent that can fetch a URL can read it without prior setup. MCP usually depends on a connection the user has already made.

## The Bottom Line

llms.txt and MCP are not substitutes. One is a briefing and the other is an interface, and a product that wants agents to both understand it and operate it needs each. The ordering is straightforward: llms.txt and Markdown page copies come first because they are cheap, public and useful to the coding agents that are already reading documentation, and MCP follows when there is something worth doing that reading alone cannot accomplish.

Expectations should be set by the evidence. llms.txt has a demonstrated role in helping agents navigate documentation and no demonstrated role in AI search visibility; most such files are never requested. MCP has enormous supply and little public data on use, and it carries real security and maintenance costs. Neither is a growth tactic on its own.

Choose llms.txt alone for a public content site with nothing for an agent to do but read. Choose MCP when agents need live, private or transactional access. For developer tools, where agents are increasingly the first evaluator, plan on both, and measure each with logs instead of assuming it works.

## Related Topics

- [llms.txt](https://metavert.io/llms-txt)
- [Model Context Protocol](https://metavert.io/model-context-protocol)
- [MCP Registry](https://metavert.io/mcp-registry)
- [Agent Experience (AX)](https://metavert.io/agent-experience)
- [AI Coding Agents](https://metavert.io/ai-coding-agents)
- [AGENTS.md](https://metavert.io/agents-md)
- [llms.txt vs robots.txt](https://metavert.io/compare/llms-txt-vs-robots-txt)
- [LightCMS](https://metavert.io/lightcms)
- [Markdown](https://metavert.io/markdown)

## Further Reading

- [The /llms.txt file proposal](https://llmstxt.org/) – Jeremy Howard, llmstxt.org
- [What is the Model Context Protocol (MCP)?](https://modelcontextprotocol.io/docs/getting-started/intro) – Model Context Protocol documentation
- [llms.txt log-file study of 137,210 domains (June 2026)](https://ahrefs.com/blog/llmstxt-study/) – Ahrefs
- [The State of Docs Traffic: A 2026 Midyear Report (July 2026)](https://www.mintlify.com/blog/state-of-docs-traffic) – Mintlify
- [Formation of the Agentic AI Foundation (December 2025)](https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation) – Linux Foundation
- [The MCP Registry](https://modelcontextprotocol.io/registry/about) – Model Context Protocol documentation
- [MCP server directory (checked October 2026)](https://www.pulsemcp.com/servers) – PulseMCP
- [MCP server directory (checked October 2026)](https://glama.ai/mcp/servers) – Glama