vaal-code 0.7.0

AI coding agent CLI — local LLMs via Ollama, AirLLM, or cloud models via API

Vaal

AI coding agent CLI. Runs local LLMs via Ollama or cloud models via API. Named after the Vaal Orb from Path of Exile.

Install

git clone https://github.com/0xE666/vaal.git
cd vaal
python install.py

That's it. vaal is now available as a command from anywhere, just like claude.

Or manually:

pip install -e .

Quick Start

# With Ollama (free, local)
vaal

# With Claude
vaal -m anthropic:claude-sonnet-4-20250514

# With OpenAI
vaal -m openai:gpt-4o

On first run with a cloud model, use /login to add your API key:

/login anthropic sk-ant-api03-your-key-here
/login openai sk-your-key-here

Or create a vaal account at e-e.tools/vaal:

/register

Features

• 45 tools — bash, file read/write/edit, git, grep, glob, web search, agents, notebooks, and more
• Subagents — spawn background agents (explore, plan, verify, code-review) that work in parallel
• 80+ slash commands — /commit, /test, /pr, /diff, /pin, /remote, /plugin, and more
• Multi-provider — Ollama, AirLLM (local 70B+ on 4GB VRAM), Anthropic, OpenAI, Groq, OpenRouter, DeepSeek, Together
• Self-repair — catches its own crashes, asks the LLM to fix the bug, patches itself
• Remote control — control vaal from a browser at e-e.tools/vaal
• Plugins — create custom tools with /plugin create or let the LLM generate them
• Tab completion — file paths auto-complete on Tab
• Persistent history — Up arrow recalls prompts across sessions

Providers

Provider	Auth	Example
Ollama	None (local)	vaal -m qwen3:8b
AirLLM	None (local, CUDA)	vaal -m airllm:Qwen/Qwen2.5-14B-Instruct
Anthropic	API key or Max/Pro OAuth	vaal -m anthropic:claude-opus-4-6
OpenAI	API key	vaal -m openai:gpt-4o
Groq	API key	vaal -m groq:llama-3.3-70b-versatile
OpenRouter	API key	vaal -m openrouter:anthropic/claude-sonnet-4
DeepSeek	API key	vaal -m deepseek:deepseek-chat
Together	API key	vaal -m together:meta-llama/Llama-3.3-70B

Key Commands

/help          — All commands
/login         — Add API key
/register      — Create vaal account
/commit        — Auto-generate commit message
/test          — Run tests, auto-fix failures
/pr            — Create pull request
/pin <file>    — Keep file in context
/remote        — Control from browser
/plugin create — Create custom tools
/effort        — Set inference level
/doctor        — System diagnostics

AirLLM — Run 70B Models Locally

Run any HuggingFace model on low-VRAM GPUs via layer-by-layer inference:

pip install vaal-code[airllm]      # install AirLLM support

vaal
/pull airllm:Qwen/Qwen2.5-14B-Instruct          # download model
/model airllm:Qwen/Qwen2.5-14B-Instruct         # switch to it

# With 4-bit compression for larger models (72B on 12GB VRAM):
/pull airllm:Qwen/Qwen2.5-72B-Instruct
/model airllm:Qwen/Qwen2.5-72B-Instruct:4bit

Set download location in ~/.vaal/config.json:

{ "airllm": { "layer_shards_saving_path": "D:/models" } }

Requirements

• Python 3.11+
• httpx, rich (auto-installed)
• Ollama (optional, for local models via Ollama)
• CUDA GPU + pip install vaal-code[airllm] (optional, for AirLLM local inference)
• Git (optional, for git commands)