nanocode is an ultra-lightweight reimplementation of Claude Code — Anthropic's official CLI agent — built from scratch by studying the decompiled source of @anthropic-ai/claude-code v2.1.88.

⚡ ~9,000 lines of code (14K total with comments · 1.8% of Claude Code's ~512K) while retaining ~85-90% of agent capability.

Disclaimer: This project is for educational, research, and technical exchange purposes only. All architectural insights are derived from the publicly available npm package. Commercial use is strictly prohibited. nanocode is not affiliated with Anthropic. If any content infringes upon rights, please contact us for immediate removal.

News

• 2026-04-01 🚀 nanocode v0.1.0 released! — 9K LOC, 15 tools, 16 slash commands. npm i -g nanocode-cli and start coding!

• 2026-03-31 💥💥💥 Claude Code source leaked! — The full TypeScript source of @anthropic-ai/claude-code v2.1.88 (~512K lines) was extracted and published from the npm package. nanocode is built from the insights gained.

Why nanocode?

Claude Code is a remarkable piece of engineering — but at 512K lines, 12MB bundled, with React/Ink UI, telemetry, bridge layers, plugin marketplace, and 90+ slash commands, it's hard to study the core patterns that actually make the agent work.

nanocode extracts the 20 critical engineering decisions that drive agent performance and implements them in clean, readable TypeScript:

Remove any one of these and overall performance degrades. nanocode keeps them all.

Claude Code vs nanocode

What's kept: Everything that affects agent reasoning quality — the loop, tools, context engineering, prompts, permissions.

What's dropped: Production wrappers — UI framework, telemetry, cloud bridge, plugin ecosystem, OAuth, 70+ niche commands.

Quick Start

Install

npm install -g nanocode-cli

Configure

nanocode works with the Anthropic API or any OpenAI-compatible provider (OpenRouter, etc.):

# Anthropic direct
export ANTHROPIC_API_KEY="sk-ant-..."

# Or OpenRouter (access to all models)
export OPENROUTER_API_KEY="sk-or-..."
export ANTHROPIC_BASE_URL="https://openrouter.ai/api"

Run

# Interactive REPL
nanocode

# One-shot mode
nanocode -p "fix the bug in auth.ts"

# With options
nanocode --model opus --thinking
nanocode --dangerously-skip-permissions -p "run all tests"
nanocode --resume <session-id>

Features

15 Built-in Tools

16 Slash Commands

/help        Show available commands
/compact     Compact conversation context (with spinner)
/clear       Clear conversation history
/model       Show or change model
/cost        Show token usage and cost
/resume      Resume a previous session
/plan        Toggle plan mode (read-only)
/memory      Show NANOCODE.md / CLAUDE.md content
/config      Show current configuration
/status      Show session status
/skills      List available skills
/context     Show context window usage
/init        Analyze codebase and generate NANOCODE.md
/mcp         Show MCP server configuration
/exit        Exit nanocode (also: /quit)

Context Engineering

• 3-layer compression: Auto-compact (summarize old messages) → micro-compact (truncate large tool results) → post-compact file restoration (re-inject top 5 recent files)
• Prompt caching: Static system prompt blocks get cache_control: ephemeral — saves ~80% on cache hits
• Token estimation: 4 chars/token for text, 2 chars/token for JSON — fast enough for real-time budget checks
• Compact threshold: contextWindow - maxOutputTokens - 13,000 — triggers before overflow

Session Persistence

• Sessions stored as JSONL in ~/.nanocode/sessions/
• Resume with /resume or --resume <id>
• Messages persisted as they're created (lazy session initialization — no empty sessions)

Skills System

Place skills in .nanocode/skills/ or .claude/skills/:

---
name: my-skill
description: Does something useful
user-invocable: true
context: inline
---

Your skill prompt here. $ARGUMENTS will be replaced.

MCP Support

Configure MCP servers in .nanocode/settings.json:

{
  "mcpServers": {
    "my-server": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path"]
    }
  }
}

Multi-line Input

• Type \ at end of line → continuation prompt …
• Option+Enter (macOS) / Alt+Enter → newline within input

Configuration

nanocode supports both .nanocode/ and .claude/ config directories (.nanocode/ takes priority):

Hierarchical loading: walks from cwd up to filesystem root, merging all found files.

Architecture

nanocode/src/          9,147 LOC (14,317 total) across 55 files
├── cli.ts             CLI entry + REPL (readline, suggestions, multi-line)
├── headless.ts        SDK/programmatic mode
├── core/              Agent loop, API client, streaming executor, errors, types
├── prompt/            System prompt, compact prompt, agent prompts, cache boundary
├── context/           Compaction, post-compact, token counting, memory, git, sessions
├── tools/             15 tools + registry
├── skills/            Skill discovery, loading, execution
├── files/             File history, LRU cache, atomic writes
├── permissions/       Permission engine, rules, path validation, modes
├── mcp/               MCP stdio client, config, types
├── commands/          16 slash commands
└── utils/             Cost tracking, ANSI formatting, spinner, process management

The Core Loop

Development

git clone https://github.com/anthropics/nanocode.git
cd nanocode
npm install
npm run build       # compile TypeScript
npm run dev         # watch mode
npm test            # run 559 tests
npm start           # run the CLI

Roadmap & Contributing

PRs welcome! The codebase is intentionally small and readable.

• More providers — Native support for OpenAI, Gemini, DeepSeek, Ollama, and other LLM providers (beyond OpenRouter passthrough)
• More tools — Git commit, PR review, code search with LSP, image generation
• More slash commands — /diff, /export, /doctor, /copy, /vim
• Streaming improvements — Server-sent events, web UI mode
• Better MCP — SSE/WebSocket transports, tool filtering, auto-reconnect
• Agent teams — Multi-agent coordination with shared task boards
• Plugin system — Installable tool/command packs
• Voice mode — Push-to-talk input via microphone
• Web UI — Browser-based interface alongside the terminal REPL
• i18n — Multilingual system prompts and UI

Pick an item and open a PR — or suggest new ideas in Issues.

License

MIT License — see LICENSE.

This project is for educational and research purposes only. Commercial use is prohibited. nanocode is not affiliated with or endorsed by Anthropic.

_{Built by studying Claude Code v2.1.88 · 1.8% of the code, ~90% of the capability}