llmstxt-mcp — MCP LLMS-TXT Documentation Server

Installed from PyPI as llmstxt-mcp-server; the CLI command is mcpdoc.

A maintained fork of langchain-ai/mcpdoc, ported to MCP Python SDK v2.

Upstream mcpdoc targets SDK v1 and declares an unbounded mcp[cli]>=1.4.1. Now that MCP Python SDK v2.0.0 (the 2026-07-28 spec release) is on PyPI, a fresh uvx --from mcpdoc mcpdoc ... resolves mcp 2.x and fails immediately — SDK v2 removed mcp.server.fastmcp and renamed FastMCP to MCPServer.

This fork ports the server to MCPServer, pins mcp>=2.0.0,<3, and adds the streamable-http transport. Original work © LangChain, Inc., MIT licensed; this fork keeps the same license. See Differences from upstream.

The PyPI distribution is llmstxt-mcp-server (PyPI rejects llmstxt-mcp as too similar to the unrelated llms-txt-mcp project). The import package and CLI command remain mcpdoc.

Still on SDK v1? Keep using upstream mcpdoc and pin mcp<2 yourself.

Overview

llms.txt is a website index for LLMs, providing background information, guidance, and links to detailed markdown files. IDEs like Cursor and Windsurf or apps like Claude Code/Desktop can use llms.txt to retrieve context for tasks. However, these apps use different built-in tools to read and process files like llms.txt. The retrieval process can be opaque, and there is not always a way to audit the tool calls or the context returned.

MCP offers a way for developers to have full control over tools used by these applications. Here, we create an open source MCP server to provide MCP host applications (e.g., Cursor, Windsurf, Claude Code/Desktop) with (1) a user-defined list of llms.txt files and (2) a simple fetch_docs tool read URLs within any of the provided llms.txt files. This allows the user to audit each tool call as well as the context returned.

llms-txt

You can find llms.txt files for langgraph and langchain here:

Library	llms.txt
LangGraph Python	https://langchain-ai.github.io/langgraph/llms.txt
LangGraph JS	https://langchain-ai.github.io/langgraphjs/llms.txt
LangChain Python	https://python.langchain.com/llms.txt
LangChain JS	https://js.langchain.com/llms.txt

Quickstart

Install uv

• Please see official uv docs for other ways to install uv.

curl -LsSf https://astral.sh/uv/install.sh | sh

Choose an llms.txt file to use.

• For example, here's the LangGraph llms.txt file.

Note: Security and Domain Access Control

For security reasons, mcpdoc implements strict domain access controls:

1. Remote llms.txt files: When you specify a remote llms.txt URL (e.g., https://langchain-ai.github.io/langgraph/llms.txt), mcpdoc automatically adds only that specific domain (langchain-ai.github.io) to the allowed domains list. This means the tool can only fetch documentation from URLs on that domain.

2. Local llms.txt files: When using a local file, NO domains are automatically added to the allowed list. You MUST explicitly specify which domains to allow using the --allowed-domains parameter.

3. Adding additional domains: To allow fetching from domains beyond those automatically included:

   • Use --allowed-domains domain1.com domain2.com to add specific domains
   • Use --allowed-domains '*' to allow all domains (use with caution)

This security measure prevents unauthorized access to domains not explicitly approved by the user, ensuring that documentation can only be retrieved from trusted sources.

(Optional) Test the MCP server locally with your llms.txt file(s) of choice:

uvx --from llmstxt-mcp-server mcpdoc \
    --urls "LangGraph:https://langchain-ai.github.io/langgraph/llms.txt" "LangChain:https://python.langchain.com/llms.txt" \
    --transport sse \
    --port 8082 \
    --host localhost

• This should run at: http://localhost:8082

• Run MCP inspector and connect to the running server:

npx @modelcontextprotocol/inspector

• Here, you can test the tool calls.

Connect to Cursor

• Open Cursor Settings and MCP tab.
• This will open the ~/.cursor/mcp.json file.

• Paste the following into the file (we use the langgraph-docs-mcp name and link to the LangGraph llms.txt).

{
  "mcpServers": {
    "langgraph-docs-mcp": {
      "command": "uvx",
      "args": [
        "--from",
        "llmstxt-mcp-server",
        "mcpdoc",
        "--urls",
        "LangGraph:https://langchain-ai.github.io/langgraph/llms.txt LangChain:https://python.langchain.com/llms.txt",
        "--transport",
        "stdio"
      ]
    }
  }
}

• Confirm that the server is running in your Cursor Settings/MCP tab.
• Best practice is to then update Cursor Global (User) rules.
• Open Cursor Settings/Rules and update User Rules with the following (or similar):

for ANY question about LangGraph, use the langgraph-docs-mcp server to help answer -- 
+ call list_doc_sources tool to get the available llms.txt file
+ call fetch_docs tool to read it
+ reflect on the urls in llms.txt 
+ reflect on the input question 
+ call fetch_docs on any urls relevant to the question
+ use this to answer the question

• CMD+L (on Mac) to open chat.
• Ensure agent is selected.

Then, try an example prompt, such as:

what are types of memory in LangGraph?

Connect to Windsurf

• Open Cascade with CMD+L (on Mac).
• Click Configure MCP to open the config file, ~/.codeium/windsurf/mcp_config.json.
• Update with langgraph-docs-mcp as noted above.

• Update Windsurf Rules/Global rules with the following (or similar):

for ANY question about LangGraph, use the langgraph-docs-mcp server to help answer -- 
+ call list_doc_sources tool to get the available llms.txt file
+ call fetch_docs tool to read it
+ reflect on the urls in llms.txt 
+ reflect on the input question 
+ call fetch_docs on any urls relevant to the question

Then, try the example prompt:

• It will perform your tool calls.

Connect to Claude Desktop

• Open Settings/Developer to update ~/Library/Application\ Support/Claude/claude_desktop_config.json.
• Update with langgraph-docs-mcp as noted above.
• Restart Claude Desktop app.

[!Note] If you run into issues with Python version incompatibility when trying to add MCPDoc tools to Claude Desktop, you can explicitly specify the filepath to python executable in the uvx command.

Example configuration

{
  "mcpServers": {
    "langgraph-docs-mcp": {
      "command": "uvx",
      "args": [
        "--python",
        "/path/to/python",
        "--from",
        "llmstxt-mcp-server",
        "mcpdoc",
        "--urls",
        "LangGraph:https://langchain-ai.github.io/langgraph/llms.txt",
        "--transport",
        "stdio"
      ]
    }
  }
}

[!Note] Currently (3/21/25) it appears that Claude Desktop does not support rules for global rules, so appending the following to your prompt.

<rules>
for ANY question about LangGraph, use the langgraph-docs-mcp server to help answer -- 
+ call list_doc_sources tool to get the available llms.txt file
+ call fetch_docs tool to read it
+ reflect on the urls in llms.txt 
+ reflect on the input question 
+ call fetch_docs on any urls relevant to the question
</rules>

• You will see your tools visible in the bottom right of your chat input.

Then, try the example prompt:

• It will ask to approve tool calls as it processes your request.

Connect to Claude Code

• In a terminal after installing Claude Code, run this command to add the MCP server to your project:

claude mcp add-json langgraph-docs '{"type":"stdio","command":"uvx" ,"args":["--from", "llmstxt-mcp-server", "mcpdoc", "--urls", "langgraph:https://langchain-ai.github.io/langgraph/llms.txt", "LangChain:https://python.langchain.com/llms.txt"]}' -s local

• You will see ~/.claude.json updated.
• Test by launching Claude Code and running to view your tools:

$ Claude
$ /mcp 

[!Note] Currently (3/21/25) it appears that Claude Code does not support rules for global rules, so appending the following to your prompt.

<rules>
for ANY question about LangGraph, use the langgraph-docs-mcp server to help answer -- 
+ call list_doc_sources tool to get the available llms.txt file
+ call fetch_docs tool to read it
+ reflect on the urls in llms.txt 
+ reflect on the input question 
+ call fetch_docs on any urls relevant to the question
</rules>

Then, try the example prompt:

• It will ask to approve tool calls.

Command-line Interface

The mcpdoc command provides a simple CLI for launching the documentation server.

You can specify documentation sources in three ways, and these can be combined:

1. Using a YAML config file:

• This will load the LangGraph Python documentation from the sample_config.yaml file in this repo.

mcpdoc --yaml sample_config.yaml

2. Using a JSON config file:

• This will load the LangGraph Python documentation from the sample_config.json file in this repo.

mcpdoc --json sample_config.json

3. Directly specifying llms.txt URLs with optional names:

• URLs can be specified either as plain URLs or with optional names using the format name:url.
• You can specify multiple URLs by using the --urls parameter multiple times.
• This is how we loaded llms.txt for the MCP server above.

mcpdoc --urls LangGraph:https://langchain-ai.github.io/langgraph/llms.txt --urls LangChain:https://python.langchain.com/llms.txt

You can also combine these methods to merge documentation sources:

mcpdoc --yaml sample_config.yaml --json sample_config.json --urls LangGraph:https://langchain-ai.github.io/langgraph/llms.txt --urls LangChain:https://python.langchain.com/llms.txt

Additional Options

• --follow-redirects: Follow HTTP redirects (defaults to False)
• --timeout SECONDS: HTTP request timeout in seconds (defaults to 10.0)
• --log-level LEVEL: Server log level — DEBUG, INFO, WARNING, ERROR, CRITICAL (defaults to INFO)

Example with additional options:

mcpdoc --yaml sample_config.yaml --follow-redirects --timeout 15

This will load the LangGraph Python documentation with a 15-second timeout and follow any HTTP redirects if necessary.

Transports

--transport accepts three values:

Transport	When to use
stdio (default)	Local MCP clients — Cursor, Windsurf, Claude Desktop, Claude Code
streamable-http	Recommended for remote hosting. Runs stateless per the 2026-07-28 spec, so replicas sit behind a plain round-robin load balancer
sse	Legacy HTTP transport; kept for compatibility with older clients

--host and --port apply to sse and streamable-http (they are ignored for stdio).

mcpdoc --yaml sample_config.yaml --transport streamable-http --host 0.0.0.0 --port 8000

Differences from upstream

	upstream mcpdoc	llmstxt-mcp-server
MCP Python SDK	v1 (FastMCP), unbounded mcp[cli]>=1.4.1	v2 only (MCPServer), mcp>=2.0.0,<3
Transports	stdio, sse	stdio, sse, streamable-http (stateless)
--log-level	parsed but ignored (hardcoded to INFO)	actually applied to the server
Tests	pure helper functions only	plus in-process client tests that build and call the real server
CI	tests the locked SDK version only	plus a job that resolves the newest allowed mcp and warns on a new SDK major

The import package and CLI command are still named mcpdoc, so migrating from upstream only means changing --from mcpdoc to --from llmstxt-mcp-server in your MCP client config.

Configuration Format

Both YAML and JSON configuration files should contain a list of documentation sources.

Each source must include an llms_txt URL and can optionally include a name:

YAML Configuration Example (sample_config.yaml)

# Sample configuration for mcp-mcpdoc server
# Each entry must have a llms_txt URL and optionally a name
- name: LangGraph Python
  llms_txt: https://langchain-ai.github.io/langgraph/llms.txt

JSON Configuration Example (sample_config.json)

[
  {
    "name": "LangGraph Python",
    "llms_txt": "https://langchain-ai.github.io/langgraph/llms.txt"
  }
]

Programmatic Usage

from mcpdoc.main import create_server

# Create a server with documentation sources
server = create_server(
    [
        {
            "name": "LangGraph Python",
            "llms_txt": "https://langchain-ai.github.io/langgraph/llms.txt",
        },
        # You can add multiple documentation sources
        # {
        #     "name": "Another Documentation",
        #     "llms_txt": "https://example.com/llms.txt",
        # },
    ],
    follow_redirects=True,
    timeout=15.0,
)

# Run the server. Under SDK v2, host/port belong to run(), not to create_server().
server.run(transport="stdio")
# or, for remote hosting:
# server.run(transport="streamable-http", host="0.0.0.0", port=8000, stateless_http=True)

create_server returns an mcp.server.mcpserver.MCPServer, so you can also drive it in-process with the SDK's test client — no subprocess needed:

from mcp.client import Client

async with Client(server) as client:
    result = await client.call_tool("list_doc_sources", {})