CodeWhale — Open-Weight AI Coding Agent for the Terminal

Introduction

CodeWhale is a terminal coding agent written in Rust that targets open-source and open-weight language models. It provides an interactive CLI for code generation, editing, and explanation without requiring proprietary API keys, making it suitable for air-gapped or privacy-sensitive environments.

What CodeWhale Does

• Provides an interactive terminal UI for AI-assisted coding tasks
• Supports multiple open-weight models including DeepSeek, Llama, and Qwen
• Reads and edits files in your project directory with contextual awareness
• Generates code, fixes bugs, and explains existing code from natural language prompts
• Runs entirely locally with no data sent to external servers

Architecture Overview

CodeWhale is a single Rust binary that communicates with local model servers via the OpenAI-compatible API format. It includes a built-in TUI for interactive sessions, a file indexer for project context, and a tool-use layer that allows the model to read, write, and search files. The Rust implementation keeps memory usage low and startup instant.

Self-Hosting & Configuration

• Requires a local model server such as Ollama, llama.cpp, or vLLM
• Install via cargo or download pre-built binaries from GitHub releases
• Configure the model endpoint and default model in ~/.codewhale/config.toml
• Set project-specific instructions via a .codewhale file in your repo root
• Supports environment variables for API endpoint and model selection

Key Features

• Pure Rust implementation for fast startup and low resource usage
• Works with any OpenAI-compatible local model server
• Built-in TUI with syntax highlighting and diff previews
• No telemetry or external data collection
• Extensible tool system for custom file operations

Comparison with Similar Tools

• Claude Code — proprietary, requires Anthropic API access; CodeWhale runs fully local
• Aider — Python-based, supports many providers; CodeWhale is Rust-native and local-first
• OpenCode — similar local-first approach; CodeWhale focuses specifically on open-weight models
• Continue — IDE extension model; CodeWhale is terminal-native

FAQ

Q: What models does CodeWhale support? A: Any model served via an OpenAI-compatible API endpoint, including DeepSeek, Llama, Qwen, and Mistral variants.

Q: Does it require a GPU? A: CodeWhale itself does not, but the underlying model server benefits from GPU acceleration for faster inference.

Q: Can I use it with remote APIs? A: Yes. Point it at any OpenAI-compatible endpoint, local or remote.

Q: How does it handle large codebases? A: It indexes project files and selectively includes relevant context in prompts, keeping token usage efficient.

Sources

• https://github.com/Hmbown/CodeWhale