Claude Code

Claude Code is an independent, research-oriented source-built reverse engineering and reconstruction project of Anthropic's official Claude Code CLI.

This project aims to provide a fully runnable, buildable, and debuggable version of the code directly from source code—liberating developers from closed-source binaries. In addition, it extends the experience with native multi-provider routing, adapters, and advanced developer utilities.

Disclaimer: This project is not affiliated with, endorsed by, or sponsored by Anthropic PBC. The original Claude Code product is proprietary. This project reconstructs and extends its behaviors strictly for research and self-hosted development environments. Please read LICENSE.md before distributing or using it in an enterprise setting.

Project Positioning

| Dimension | What Claude Code Provides | | --- | --- | | Original Fidelity | A reconstructed CLI conforming to the terminal UX, tools, and extension points of the official Claude Code. | | Build & Debug | A modular Bun/TypeScript codebase tree that can be typechecked, tested, and modified locally using bun run dev. | | Enterprise Capabilities | Remote session bridging, MCP, plugins, skills, agents/supervisor, voice, session memory, and LSP integrations—without forcing all workflows to run through Anthropic's hosted-only services. | | Our Superpowers | Declarative multi-provider routing (providers.json, /model), provider-specific adapters, and custom developer utilities (preload, codeindex, session). |

This is a community rebuild for software engineers who require transparency and provider choice—not an official distribution from Anthropic.

Features

Claude Code is an AI coding assistant that runs directly in your terminal. It inspects and edits local codebases, executes shell commands, switches model providers, and coordinates complex, long-running agent workflows through custom commands, supervisors, plugins, and skills.

Key Highlights:

• Multi-Provider AI Routing — Supports Anthropic, OpenAI, Google Gemini, OpenRouter, Ollama, GitHub Copilot, and other OpenAI-compatible endpoints.
• Runtime Model Swapping — Switch models or providers on-the-fly using the /model slash command.
• Tool-Driven Workflows — Automatically read, search, edit, and write files, execute shell commands, query LSPs, browser automation, and run MCP tools.
• Plugin Hooks System — Intercept and hook into prompts, shell executions, tool invocations, and file editing actions.
• Dynamic Skills — Load capabilities from bundled sources and custom project-level .claude/skills/ directories.
• Agents & Supervisor — Orchestrate deep research, coding, and multi-agent coordination.
• Durable Agent Runtime & Orchestrator (PLAN I) — A highly resilient, 100% offline-friendly agent runtime featuring automated checkpoint recovery and interactive user approvals for sensitive commands.
• Scheduled Tasks — Create one-shot or recurring automation tasks via the interactive /task panel, backed by durable storage (.claude/scheduled_tasks.json).
• Sessions & Bridge — Save context, pause/resume tasks across sessions, and enable remote workspace collaboration.

Quick Start

Global Installation

npm install -g @jonusnattapong/claudecode

npm install -g installs the launcher, but claudevil still requires Bun to be installed on your machine at runtime.

Or:

bun install -g @jonusnattapong/claudecode

Run the assistant in any of your project directories:

claudevil

Running From Source

git clone https://github.com/JonusNattapong/claudecode.git
cd claudecode
bun install
bun run build
bun run start

System Requirements

• Bun 1.3 or higher for local development and for the globally installed claudevil launcher.
• At least one API key, such as ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY, or other supported providers.
• Windows, macOS, Linux, or WSL2.

Provider Configuration

Configure keys in your shell or inside a local .env file at the root of the project:

export ANTHROPIC_API_KEY=sk-ant-...
export OPENAI_API_KEY=sk-...
export GOOGLE_API_KEY=...
export OPENROUTER_API_KEY=sk-or-...
export OLLAMA_HOST=http://localhost:11434

To switch models or providers inside a live session:

/model
/model list
/model openai/gpt-4o
/model google/gemini-2.5-pro

Provider Configuration Overview: docs/providers.html

Frequently Used Commands

/model      Switch models or providers at runtime
/status     Show provider session and context status
/doctor     Run diagnostic tests and auto-fixes
/context    Inspect detail of the active context window
/compact    Compress and prune conversation history
/mcp        Manage Model Context Protocol (MCP) servers
/plugin     Manage plugins and lifecycle hooks
/bridge     Configure bridge mode for remote collaboration
/agent      Manage background agent workflows (run, status, trace, approvals, report)
/daemon     Open an interactive dashboard for autonomous background daemons
/task       Create Scheduled Tasks or manage the autonomous task queue

Type / in the CLI prompt at any time to see the autocomplete list of all supported commands.

Scheduled Tasks

The Scheduled Tasks system is powered by an interactive form in /task—eliminating the need to memorize complex cron expressions. Type /task with no arguments, fill in the fields, and submit.

| What You Do | What Claude Code Does | | --- | --- | | /task | Opens the interactive form to create a Scheduled Task | | Select Daily at 09:00 | Creates a daily recurring task | | Select Weekdays at 09:00 | Creates a weekday cron schedule (e.g., 0 9 * * 1-5) | | Select In N minutes with value 10 | Creates a one-shot reminder relative to your local timezone | | Select Custom cron | Manually enter standard 5-field cron syntax | | /task scheduled | Directly opens the scheduling form | | /task list | Lists all active tasks in the autonomous queue |

Details:

• Uses standard 5-field cron according to your local machine's timezone: minute hour day-of-month month day-of-week
• Durable tasks are permanently saved to .claude/scheduled_tasks.json and persist across sessions.
• Session-only tasks are stored in volatile memory and run strictly during the active session.
• Recurring tasks auto-expire after 30 days (except system-generated permanent tasks).
• One-shot tasks are automatically deleted after they fire.
• Natural language time scheduling remains supported behind the scenes via CronCreate, CronList, and CronDelete tools when selected by the model.

Examples:

/task
Name: Server Check
Schedule: Daily
Time: 20:00
Prompt: Verify the status of local servers
Storage: Durable

/task
Name: Commit Reminder
Schedule: In N minutes
Delay: 10
Prompt: Don't forget to commit the code
Storage: Session-only

Development

bun run dev              # Starts development mode with hot-reload
bun run start            # Runs the CLI from source code
bun run build            # Compiles and builds the bundle into dist/
bun test                 # Runs the test suite
bun x tsc --noEmit       # Runs TypeScript typechecks
bun run lint:check       # Checks Biome lint rules
bun run format:check     # Checks Biome formatting
bun run check:ci         # Runs Biome CI validation

In-project Developer Utilities:

bun run preload <module>     # Code Preloader: Preload module contexts before editing
bun run session <command>    # Session Bridge: Save, list, or restore terminal context across sessions
bun run codeindex <command>  # CodeIndex: Index and fuzzy-query the codebase
bun run codegraph            # CodeGraph: Generate module dependency graphs
bun run ast-grep -- <args>   # ast-grep: Perform structural AST-based search/rewriting

Project Structure

src/
├── main.tsx              # Ink React terminal UI bootstrap & main loop
├── query.ts              # Core query processing and system prompts
├── QueryEngine.ts        # Query orchestrator (caching, deduplication, rate limits)
├── agentRuntime/         # Agent orchestration, persistent run stores, and tool gateways
├── commands/             # Slash command implementations
├── tools/                # Built-in developer tools
├── services/
│   ├── ai/               # ProviderManager, adapters, normalizers, and providers.json
│   ├── mcp/              # Model Context Protocol clients
│   ├── plugins/          # Plugin lifecycle hooks and interceptors
│   ├── tools/            # Tool execution service
│   ├── lsp/              # Language Server Protocol integration
│   ├── Supervisor/       # Autonomous agent supervisor
│   └── SessionMemory/    # Persistent session memory
├── skills/               # Dynamic project-level skills loader
├── cli/                  # Terminal UI contexts
├── components/           # Terminal UI rendering components
├── bridge/               # WebSocket bridge for remote pairing
├── coordinator/          # Multi-agent coordinator and worker setup
├── keybindings/          # Keyboard shortcut mappings
├── state/                # Reactive store implementations
└── vim/                  # Vim-like keyboard navigation mode

Architecture

Terminal UI
  -> Command Registry & Keybindings Layer
  -> Provider Manager & AI Adapters
  -> Query Engine & Streaming Loops
  -> Tool Executor Service
  -> Plugin hooks, MCP client, LSP integration, agents, session memory, bridge

Documentation

• Installation
• Quick Start
• Configuration
• AI Providers
• Models
• Commands
• Tools
• Plugins
• Skills
• Architecture
• Permission Model
• Bridge Mode
• SearXNG Search
• Troubleshooting
• Evals

Debugging

DEBUG=1 bun run src/main.tsx
DEBUG=provider:anthropic bun run src/main.tsx

Platform Notes

Windows

Remove-Item -Recurse -Force node_modules
bun install
bun run dev

• Precompiled ripgrep for Windows is bundled under src/utils/vendor/ripgrep/x64-win32/rg.exe to guarantee lightning-fast filesystem searches.

Contributing

Please refer to CONTRIBUTING.md, CODE_OF_CONDUCT.md, and SECURITY.md.

Changelog

CHANGELOG.md

License

LICENSE.md