For AI agentsA Markdown index of these docs is at /llms.txt. Every page is also Markdown: add .md to its address.

Docs / Set up with your agent

LLMGRAPH for Claude Code

Add the LLMGRAPH MCP server to Claude Code or install the plugin on Windows: the config file, the cmd /c npx form, and how to check it connected.

Add LLMGRAPH to Claude Code with one command, or install the plugin that adds the server and teaches your agent when to ask. Claude Code is tested by us: we ran it against the server on Windows 11, where it starts npx directly.

1Add the server to Claude Code

Add tested by us a command, or .mcp.json in the project root (project scope)

$claude mcp add --scope project llmgraph -- npx -y llmgraph-mcp

Or as a file entry:

json{ "mcpServers": { "llmgraph": { "command": "npx", "args": [ "-y", "llmgraph-mcp" ] } } }

On Windows 11, Claude Code starts npx directly, so no cmd /c wrapper is needed.

Or install the plugin

The plugin adds the server and a short skill that tells your agent when to ask which question. From the Claude Code plugin marketplace:

In a shell:

$claude plugin marketplace add brainfoods-studio/llmgraph-agent-plugin
$claude plugin install llmgraph@llmgraph

Or inside a Claude Code session:

/plugin marketplace add brainfoods-studio/llmgraph-agent-plugin
/plugin install llmgraph@llmgraph

The plugin also adds the stop check, which holds Claude Code once when its edits left callers certain to break. More in The LLMGRAPH plugin.

**"Pending approval" in claude mcp list?** Claude Code shows this for a server added in project scope until you approve it in an interactive session (run claude). The line is not a sign that LLMGRAPH is off.

2Check that it connected

Check tested by us (your agent lists them as tools)

  • Run claude mcp list, or type /mcp inside Claude Code.
  • A project-scope server from .mcp.json first shows 'Pending approval' until you approve it in an interactive session.

You should see ✔ Connected.

3Sign in once

The first time your agent asks a question, LLMGRAPH answers with a sign-in link and a 6-character code. In a terminal you can run this instead:

$npx -y llmgraph-mcp login

The link works for 10 minutes. Open the full link, including everything after the #, sign in with Google, compare the code on the page with the one your agent shows, and approve this device. Approving your first device starts your 7-day trial. The full steps are in Set up with your agent.

4Make sure Claude Code uses it

Adding a server doesn't make an agent use it.

Add a line to your agent's instructions file, CLAUDE.md:

CLAUDE.mdBefore changing code, ask the llmgraph MCP tools who_calls and what_breaks_if_changed; after editing, run check_my_changes; read each answer's label and its lines about what it cannot see.

We can't promise that an agent asks on its own. Every answer starts with llmgraph · : look for it where Claude Code shows what it asked.

5The five questions

QuestionWhat it answers
who_callsEvery place that calls this function, plus calls we can't see.
what_breaks_if_changedEverything that depends on this, directly or indirectly, and what we can't see.
check_my_changesAfter editing: callers you didn't update, and anything we can't see.
what_is_unknownWhere our view is blind (dynamic calls, unknown types, unreadable files) and why.
explainWhy we said this: the evidence behind one answer line.

Other agents

For agentsthis page as Markdown/docs/agents/claude-code.mdllms.txt