Mapping Claude Code: A Month-Long Deep Dive into 512K Lines of AI Agent Source

Source‑map leak and code size

On 31 Mar 2026 Chaofan Shou discovered a .map file in the npm package anthropic/claude-code. The source map allowed reconstruction of the full TypeScript codebase: 1 884 files, more than 512 000 lines of code. The largest file is screens/REPL.tsx at 875 KB (≈25 000 lines). The next largest file is main.tsx at 785 KB.

Code composition

1 332 .ts files + 552 .tsx files = 1 884 files

512 000+ lines of code

Utilities in src/utils/: 329 files, with a single utils/hooks.ts file of 156 KB

80+ slash commands in src/commands/ 85 React hook files, 144 UI component files

Harness vs. model interaction

Only about 5 % of the code directly calls the model API; the remaining ~95 % implements the “harness” that handles compression, validation, permission checks, caching, hot‑updates, and other runtime concerns.

Technical choices

Runtime : Bun – --version takes 12 ms vs Node’s 50 ms, keeping CLI start‑up within a 200 ms latency budget.

Language : TypeScript (strict) – prevents technical debt in a half‑million‑line codebase.

Terminal UI : React + a forked Ink ( src/ink/, 48 files) – provides click events, focus detection, hyperlinks, and incremental updates not available in the npm Ink package.

CLI parsing : Commander.js with extra typings – full type inference for commands.

Schema validation : Zod v4 – validates tool inputs, hook configs, and MCP schemas.

Code search : ripgrep – used by GrepTool for the fastest regex search.

Protocol : MCP + LSP – external tools use MCP, code‑intelligence uses LSP.

Telemetry : OpenTelemetry + gRPC – lazy‑loaded (~400 KB OTel, ~700 KB gRPC) to keep binary size low.

Feature flags : GrowthBook + custom feature() – compile‑time gating, stripped from the final build.

Auth : OAuth 2.0 + JWT + macOS Keychain – parallel reads spawning plutil.

Architecture layers

Brainstem (~5 %) : query.ts, QueryEngine.ts, Tool.ts, tools.ts, commands.ts – core loop of 16 steps, only one step invokes the model.

Core services (~15 %) : services/, coordinator/, tasks/, skills/, memdir/ – the “nervous system”.

Interaction layer (~30 %) : screens/, components/, hooks/, ink/, bridge/, keybindings/, vim/ – UI skin.

Support layer (~50 %) : utils/, bootstrap/, entrypoints/, migrations/, constants/ – utility functions.

Comparison with other AI coding agents

Claude Code : TypeScript, Bun runtime, React + Ink TUI, AsyncGenerator agent loop, 43+ built‑in tools plus MCP, Swarms + Coordinator for multi‑agent orchestration, >512 k lines.

Cursor Agent : TypeScript, Electron + Node runtime, web UI, no multi‑agent support, ~20 tools, code size not disclosed.

Aider : Python, CPython runtime, pure terminal UI, synchronous loop, ~10 tools, ~30 k lines.

OpenHands : Python, Docker runtime, web UI, event‑driven loop, ~20 tools, ~100 k lines.

Key judgments

Harness engineering dominates : A LangChain benchmark shows that swapping only the harness moves a model from rank 30 to rank 5, indicating that future agent competitiveness will depend more on harness quality than on the underlying model.

Reusable patterns : At least 20 patterns can be copied directly, including an AsyncGenerator‑driven loop, buildTool() fail‑closed defaults, a four‑stage context compression pipeline, typed QueryDeps, transition tracking, configuration snapshots, tool‑list partitioning, and a 34‑line minimal store.

Industrial‑grade multi‑agent orchestration : Seven TaskType s, task‑ID prefixes for operational friendliness, Unix Domain Socket communication (~50 µs RTT vs HTTP ~500 µs), and a Coordinator pattern that isolates control‑plane tools.

Fast‑path startup

The CLI command claude --version runs in 12 ms on the author’s machine, whereas node --version takes 50 ms. The speed comes from a “Fast Path” implementation in the first 30 lines of entrypoints/cli.tsx and a parallel pre‑fetch in the top of main.tsx.