Claude Code Boot Sequence: How a CLI Becomes a Runtime Host

Claude Code is not a CLI that wraps an API call. It is a runtime host that happens to present as a CLI. The distinction shows up immediately in how it boots.

The boot file hierarchy

The startup chain spans five files:

src/
├── cli.tsx
├── main.tsx
├── setup.ts
├── replLauncher.tsx
├── entrypoints/
│   └── init.ts

cli.tsx is the real entry point, not main.tsx. It implements waterfall fast-paths: --version returns before a single module is loaded. Each subcommand (daemon, bridge, ps, logs) loads only the modules it needs.

main.tsx at 4,683 lines handles everything else: auth strategies, permission init, MCP config, plugin loading, AppState creation, and Commander CLI definitions.

The annotated startup timeline

Side-effect hoisting is the most non-obvious optimization in the boot sequence. Two reads — MDM (managed device policy) and keychain (API key) — fire between import statements, running in parallel with the ~135ms of module evaluation time:

T+0ms    cli.tsx entry point
T+2ms    startMdmRawRead() and startKeychainPrefetch() fire (between imports)
T+137ms  module evaluation complete — MDM and keychain reads already done
T+150ms  API preconnection HEAD request fires (puts TCP+TLS into keep-alive pool)
T+250ms  REPL renders, buffered stdin input consumed

By the time preAction executes, the prefetches have already completed. The API preconnection saves 100–200ms on the first real model call by placing a connection into Bun's keep-alive pool before it is needed.

Stdin buffering at T+0ms means keystrokes typed during startup are not lost — they are consumed when the REPL mounts.

Trust is a boot problem, not just a tool problem

Trust checks precede most runtime assembly. Specifically:

• git-backed system context is only prefetched when the directory is already trusted or the session is non-interactive
• assistant mode is refused in untrusted directories
• some setup work is deferred until after setup screens and trust checks complete

The consequence: a user in an untrusted directory gets a constrained session, not a full one with tools enabled. Trust governs which session gets assembled, not just which tools execute within it.

Special launch modes normalize into the common runtime

Launch modes like cc:// deep links, claude assistant, and claude ssh <host> are rewritten in argv early in cli.tsx. Most land inside the full interactive runtime with its approval dialogs, streaming UX, and session resume.

This is a deliberate product choice over isolated per-mode utilities. One runtime surface is maintained and tested; edge cases do not drift into separate code paths.

The action handler is the composition root

The main .action(...) block in main.tsx is where a generic session becomes a specific one. Everything assembled here — permission mode, MCP config, agent definitions, worktree mode, remote and Chrome integrations — is passed to replLauncher as props. The REPL does not independently discover its capabilities; they are handed to it at mount.

Same concept in three languages

async function main() {
installSignalHandlers();
normalizeSpecialLaunchArgs(process.argv);

const isNonInteractive =
  hasPrintFlag(process.argv) ||
  hasInitOnlyFlag(process.argv) ||
  hasSdkUrl(process.argv) ||
  !process.stdout.isTTY;

setIsInteractive(!isNonInteractive);
initializeEntrypoint(isNonInteractive);
eagerLoadSettings();

await init();
runMigrations();

const program = buildCommanderProgram();
await program.parseAsync(process.argv);
}

program.action(async (prompt, options) => {
const permissionMode = resolvePermissionMode(options);
const toolContext = await initializeToolPermissionContext(options);
const commands = await getCommands(getCwd());
const agents = await getAgentDefinitionsWithOverrides(getCwd());

await maybeShowSetupScreens();
await launchRepl({ prompt, options, permissionMode, toolContext, commands, agents });
});

What this file proves

After this first pass, main.tsx proves three architecture points:

1. Claude Code is optimized around startup responsiveness.
2. Trust and permissions are designed into the runtime from the first stage of boot.
3. The CLI is a session host that composes capabilities before entering the main agent loop.

Source file inventory: Entrypoints and Boot

The files below form the boot layer. These are the ones worth reading before anything else.

The question is no longer:

• how does Claude Code call the model?

It becomes:

• how does Claude Code build a safe, resumable, interactive execution runtime before the model even takes the first turn?