Vibe Code

A local AI coding assistant. Runs entirely on your machine — no API keys, no cloud.

Supports two backends: ollama (recommended, easy model management) and llama-cpp (direct GGUF, full CUDA control).

Features

• Agentic tool-call loop: reads files, writes files, edits files, runs bash commands, globs, greps
• System control: GPU stats, process management, systemd services, network info
• Qwen3/Qwen3.5 thinking mode (chain-of-thought reasoning, toggleable at runtime)
• Streaming output with Rich UI
• Session memory via /save → .vibe/memory.md
• Context usage tracking with visual bar
• Opens GUI apps and file managers directly

Requirements

• Arch Linux (setup script uses pacman)
• NVIDIA GPU with CUDA support
• Python 3.12+
• ollama (recommended) or llama-cpp-python with CUDA

Setup

Ollama (recommended)

./setup.sh            # installs ollama, pulls qwen3:14b, creates venv
./vibe.sh

Other models you can pull:

ollama pull qwen3.5:9b         # smaller, faster
ollama pull qwen3:30b-a3b      # larger MoE, needs /offload for <16GB VRAM
ollama pull qwen2.5-coder:14b  # code-specialized

Switch models in vibe/config.py:

OLLAMA_MODEL = "qwen3:14b"     # change to any pulled model

llama-cpp (direct GGUF)

./setup.sh llama_cpp  # installs CUDA, builds llama-cpp-python, downloads GGUF
./vibe.sh

Set in vibe/config.py:

BACKEND = "llama_cpp"
MODEL_PATH = ROOT_DIR / "models" / "Qwen3-8B-Q8_0.gguf"

GPU offload

If a model doesn't fit entirely in VRAM, use /offload at runtime to split layers between GPU and CPU/RAM:

/offload 35    # 35 layers on GPU, rest on CPU/RAM
/offload 0     # CPU only (no VRAM used)
/offload -1    # all layers on GPU (default)

Usage

# Launch
./vibe.sh

# Point at a project directory
./vibe.sh ~/my-project

# Verbose mode (shows tracebacks)
./vibe.sh --verbose

Slash Commands

Conversation

Command	Description
/reset	Clear conversation history
/think	Enable chain-of-thought reasoning
/nothink	Disable chain-of-thought (faster)
/tokens	Show context usage with a visual bar
/model	Show current backend and model info
/save	Summarize session to .vibe/memory.md
/memory	Print current .vibe/memory.md
/exit	Quit
/help	Show all commands

System Info

Command	Description
/sys	CPU, RAM, disk usage, uptime
/gpu	GPU utilization, VRAM usage, temperature
/net	Network interfaces and listening ports
/ps [filter]	Top processes by CPU (optional name filter)
/files [path]	Open file manager (default: cwd)

Process & Service Control

Command	Description
/kill <pid|name>	Send SIGTERM to a PID or matching processes
/service <name> [action]	systemctl control (default: status). Actions: start stop restart reload enable disable
/services [filter]	List running services

Runtime Config

Command	Description
/offload <n>	GPU layers (-1=all, 0=CPU only, N=partial)
/set temp <0.0-2.0>	Generation temperature
/set tokens <n>	Max output tokens
/set top_p <0.0-1.0>	Nucleus sampling probability
/set top_k <n>	Top-k sampling
/set repeat_penalty <n>	Repetition penalty

Configuration

vibe/config.py:

# Backend
BACKEND      = "ollama"      # "ollama" or "llama_cpp"

# Ollama
OLLAMA_HOST  = "http://localhost:11434"
OLLAMA_MODEL = "qwen3:14b"
OLLAMA_NUM_GPU = -1          # GPU layers (-1 = all, 0 = CPU only)

# llama-cpp
MODEL_PATH   = ROOT_DIR / "models" / "Qwen3-8B-Q8_0.gguf"
N_CTX        = 32768         # context window (tokens)
N_GPU_LAYERS = -1            # GPU layers (-1 = all, 0 = CPU only)

# Generation (both backends)
TEMPERATURE  = 0.6
MAX_TOKENS   = 16384
THINKING     = False         # /think to enable

Project Structure

vibe-code/
├── main.py          # REPL entry point, slash command handling
├── vibe/
│   ├── config.py    # Backend selection, model paths, generation params
│   ├── llm.py       # VibeModel — agentic loop, ollama + llama-cpp backends
│   ├── tools.py     # Tool schemas + implementations (read, write, edit, bash, glob, grep, ls)
│   ├── system.py    # System commands (gpu, ps, services, file manager, etc.)
│   └── ui.py        # Rich console UI, prompt_toolkit session, streaming renderer
├── setup.sh         # llama-cpp setup script
├── vibe.sh          # Launch wrapper
└── requirements.txt