# localcode

Fully offline AI coding environment for macOS Apple Silicon. Uses **Ollama** to serve local models with **10 terminal coding agents** — no API keys, no cloud, no costs.

## Quick Start

```bash
npm run build
npm run dev -- setup     # installs Ollama, pulls models, sets up configs
localcode                # launch with defaults
```

## Usage

```bash
localcode                          # launch with last-used TUI + model
localcode goose                    # launch with Goose
localcode claude                   # launch with Claude Code
localcode gpt-oss                  # launch with GPT-OSS model
localcode goose gpt-oss            # launch with both overrides

localcode models                   # list available models
localcode tuis                     # list available TUIs
localcode status                   # show config + Ollama health

localcode start                    # start Ollama + pull models
localcode stop                     # stop Ollama
localcode bench                    # benchmark the active chat model
localcode pipe "add type hints"    # pipe stdin through the model
```

TUI and model names are auto-detected — just type the id directly. Last-used choices are saved for next time.

## TUIs

TUIs with `ollama launch` support are installed and configured automatically by Ollama.

| TUI | Launch | Method |
|-----|--------|--------|
| Claude Code | `localcode claude` | ollama launch |
| Codex CLI | `localcode codex` | ollama launch |
| OpenCode | `localcode opencode` | ollama launch |
| Pi | `localcode pi` | ollama launch |
| Cline | `localcode cline` | ollama launch |
| Droid | `localcode droid` | ollama launch |
| OpenClaw | `localcode openclaw` | ollama launch |
| Aider | `localcode aider` | direct (env vars) |
| Goose | `localcode goose` | direct (env vars) |
| gptme | `localcode gptme` | direct (env vars) |

## Models

| Model | ID | Size | Tool Calling | Notes |
|-------|----|------|-------------|-------|
| Qwen3 Coder 30B-A3B | `qwen3-coder` | 19 GB | Yes | Best coding benchmarks (SWE-bench 69.6) |
| GLM-4.7 Flash 30B-A3B | `glm-flash` | 19 GB | Yes | Strong coding, 198K context |
| GPT-OSS 20B | `gpt-oss` | 14 GB | Yes | Lightest with tool support, good for 32GB machines |
| Qwen 2.5 Coder 32B | `qwen-32b-chat` | 20 GB | No | Dense model, no structured tool calls |
| Qwen 2.5 Coder 14B | `qwen-14b-chat` | 9 GB | No | Smallest chat model |
| Qwen 2.5 Coder 7B | `qwen-7b-chat` | 5 GB | No | Lightest chat model |
| Qwen 2.5 Coder 1.5B | `qwen-1.5b-autocomplete` | 1 GB | No | Autocomplete only |

Models without tool calling will output raw tool-call text in agents that rely on structured tool use (Goose, Pi, etc.). Use gpt-oss, qwen3-coder, or glm-flash for those agents.

## Hardware Requirements

- **Mac with Apple Silicon** (M1/M2/M3/M4)
- **32 GB RAM** recommended
- Disk space depends on models pulled

## Configuration

All stored in `~/.config/localcode/config.json` — just the active model, autocomplete model, and TUI ids. Ollama manages model storage.

## Uninstall

```bash
brew uninstall ollama          # remove Ollama + all pulled models
rm ~/.local/bin/localcode      # remove CLI wrapper
rm -rf ~/.config/localcode     # remove config
```