An agentic coding tool powered by Ollama - local, private, and fast
npm install olliecodeAgentic coding tool powered by Ollama - local, private, and fast.
Ollie is an AI-powered coding assistant that runs entirely through Ollama. Run models locally on your machine for complete privacy, or use Ollama Cloud for access to larger models without the hardware requirements.
- Ollama Powered - Works with any Ollama model, local or cloud
- Privacy First - Run locally and your code never leaves your machine
- Agentic Workflows - Ollie can read, write, and edit files autonomously
- Multiple Modes - Plan, Build, and Explore modes for different tasks
- Tool Use - File operations, grep, glob, command execution, and more
- Session Management - Persistent conversation history across sessions
- Beautiful TUI - Modern terminal interface with syntax highlighting
- Themeable - Multiple built-in themes (Ollie, Catppuccin, Dracula, Nord, Tokyo Night, Monokai)
- Bun >= 1.0.0
- Ollama running locally (or Ollama Cloud account)
``bash`
brew tap ollielabs/tap
brew install ollie
`bashInstall globally with bun
bun add -g olliecode
$3
bash
git clone https://github.com/ollielabs/olliecode.git
cd olliecode
bun install
bun dev
`$3
Download pre-built binaries from the releases page.
Recommended Models
Ollie works with any Ollama-compatible model. Here are some recommendations based on your setup.
$3
Access powerful models without local hardware requirements via Ollama Cloud:
| Model | Best For | Notes |
|-------|----------|-------|
|
| Reasoning, agentic tasks | OpenAI's open-weight models (20B/120B) |
| gemini-3-pro-preview | Complex reasoning | Google's most intelligent model |
| gemini-3-flash-preview | Fast, cost-effective | Frontier intelligence built for speed |
| deepseek-v3.2 | Reasoning, agent performance | High computational efficiency |
| deepseek-v3.1 | Thinking mode support | Hybrid thinking/non-thinking |
| kimi-k2 | Coding agent tasks | State-of-the-art MoE model |
| kimi-k2-thinking | Deep reasoning | Moonshot AI's best thinking model |
| qwen3-coder | Agentic coding | Long context, 30B/480B variants |
| qwen3-next | Efficiency, speed | Strong tool use and thinking |
| devstral-2 | Code exploration, editing | 123B, excellent for agents |
| devstral-small-2 | Code exploration | 24B, vision + tools |
| mistral-large-3 | Enterprise workloads | General-purpose multimodal MoE |
| glm-4.7 | Coding | Advanced coding capability |
| glm-4.6 | Agentic, reasoning | Strong all-rounder |
| minimax-m2.1 | Multilingual, code | Exceptional multilingual |
| minimax-m2 | Coding, agentic | High-efficiency |
| cogito-2.1 | General purpose | 671B, MIT licensed |$3
Models specifically trained for code generation and understanding:
| Model | Parameters | VRAM | Best For |
|-------|------------|------|----------|
|
qwen3-coder | 30B | ~18GB | Best local coding model |
| devstral-small-2 | 24B | ~16GB | Excellent for code tasks |
| deepseek-coder-v2 | 16B | ~10GB | Strong code completion |
| codellama | 34B | ~20GB | Meta's code-focused model |
| starcoder2 | 15B | ~10GB | Multi-language support |$3
These models support function calling, essential for Ollie's agentic features:
| Model | Parameters | VRAM | Notes |
|-------|------------|------|-------|
|
qwen3 | 32B | ~20GB | Excellent tool use support |
| qwen3 | 14B | ~9GB | Good balance of speed/quality |
| llama3.3 | 70B | ~42GB | Best open model, needs high VRAM |
| llama4-scout | 17B | ~11GB | Meta's latest, efficient |
| mistral-small | 22B | ~14GB | Good tool use, reasonable size |
| command-r | 35B | ~22GB | Cohere's tool-focused model |
| gemma3 | 27B | ~17GB | Vision support, runs on single GPU |$3
For machines with limited resources (8GB RAM or less):
| Model | Parameters | VRAM | Notes |
|-------|------------|------|-------|
|
qwen3 | 8B | ~5GB | Best lightweight option |
| qwen3 | 4B | ~3GB | Very fast, basic tasks |
| llama3.2 | 3B | ~2GB | Minimal footprint |
| phi-4 | 3.8B | ~3GB | Microsoft's efficient model |
| gemma3 | 4B | ~3GB | Google's compact model |
| ministral-3 | 3B | ~2GB | Edge deployment ready |$3
`bash
Pull a recommended model
ollama pull qwen3:14bOr for coding-specific tasks
ollama pull qwen3-coder:30bFor limited hardware
ollama pull qwen3:8b
`Usage
$3
bash
Start with default model
ollieStart with a specific model
ollie --model qwen3:14bResume a previous session
ollie --session Start in a specific mode
ollie --mode plan
`$3
Ollie has three operational modes:
- Build (default) - For implementing features and writing code
- Plan - For designing and planning before implementation
- Explore - For understanding and navigating codebases
Switch modes with
or use keyboard shortcuts.$3
Reference files directly in your prompts using
@:`
@src/index.ts what does this file do?
@package.json update the version
$3
| Key | Action |
|-----|--------|
|
| Open command menu |
| Ctrl+N | New session |
| Ctrl+L | Clear screen |
| Ctrl+C | Cancel/Exit |
| Esc | Close modals |$3
| Command | Description |
|---------|-------------|
|
/help | Show help |
| /mode | Switch mode (plan, build, explore) |
| /model | Switch model |
| /session | Session management |
| /theme | Change theme |
| /clear | Clear conversation |
| /compact | Compact conversation history |Configuration
$3
`bash
Ollama host (default: http://localhost:11434)
OLLAMA_HOST=http://localhost:11434Default model
OLLIE_MODEL=qwen3:14bDefault mode
OLLIE_MODE=build
`$3
Usage: ollie [options]Options:
-m, --model Model to use
-s, --session Resume session
--mode Starting mode (plan, build, explore)
-v, --version Show version
-h, --help Show help
$3
Ollie stores configuration in
and session data in ~/.local/share/ollie/.Development
`bash
Install dependencies
bun installRun in development mode
bun devBuild binary
bun run buildFormat code
bun run format:fixLint
bun run lint:fixType check
bun run check:types
``Contributions are welcome! Please feel free to submit a Pull Request.
MIT License - see LICENSE for details.
Built with:
- Ollama - Local LLM runtime
- OpenTUI - Terminal UI framework
- Bun - JavaScript runtime