deuscode 0.7.4

AI-powered multi-agent CLI coding assistant for local LLMs

Deus

Claude Code-style AI coding assistant that runs on your own GPU.

The problem it solves

Proprietary coding assistants charge per token — and token anxiety kills the "just let it figure it out" workflow. Every file read, every search, every iteration costs money and sends your code to someone else's server. Deus runs on a GPU you rent by the hour: analyze your entire codebase, run the agent in loops, explore freely. Flat rate. Your data stays yours.

How it works

Your prompt
    ↓
Complexity detection (simple? skip planning)
    ↓
Planner → action plan (files to read, searches to run, files to create)
    ↓
Parallel pre-loading (reads files + runs searches before agent starts)
    ↓
Agent loop (reads files, writes code, runs commands, searches web)
    ↓
Result

The planner analyzes your task and pre-loads all relevant context before the agent starts. Instead of the agent discovering it needs a file 3 turns in, it begins with everything already loaded.

Quick start

1. Install

pip install deuscode

2. Launch a GPU on RunPod

deus setup --runpod

Interactive wizard: picks model → GPU size → launches vLLM → writes your config automatically.

💡 Support Deus development by using our RunPod affiliate link: runpod.io/ref/ww1q3uhd

• ~40x cheaper than API tokens for intensive use
• Flat hourly rate — no token counting, no throttling
• Code never leaves your pod — not sent to any third-party API
• Auto-stop built in — Deus stops your pod when done, no idle charges
• Any open-source model — Qwen2.5-Coder, DeepSeek, Llama 3, and more

3. Start coding

cd your-project
deus "add Stripe payment integration"

Or launch the interactive REPL:

deus

Features

Intelligent planning

Every prompt goes through complexity detection first. Simple questions go straight to the agent. Complex tasks get a full action plan: which files to read, which docs to search, what to create, and how to validate the result. The agent starts with context already loaded.

Web search built in

Before the agent writes a line of code, the planner can queue web searches. Latest API docs, Stack Overflow answers, package changelogs — fetched and injected into context automatically. Uses DuckDuckGo by default; swap to Brave Search with an API key.

Full tool suite

• Read/write files — diff preview and confirmation before any write
• Bash commands — explicit confirmation before execution
• Web search — DuckDuckGo (default) or Brave Search
• Repo-map — scans your codebase structure, passed to every prompt

Model management

deus model list                    # see downloaded models on your pod
deus model download --size small   # pick and download a model

Coding models:

Model	VRAM	Notes
Qwen2.5-Coder-1.5B	4 GB	Tiny, any GPU
Qwen2.5-Coder-3B	8 GB	Small but capable
Qwen2.5-Coder-7B	16 GB	Fast, cheap, good
Qwen2.5-Coder-14B	28 GB	Best mid-size
DeepSeek-Coder-V2-Lite	32 GB	MoE, strong for size
Qwen2.5-Coder-32B	64 GB	Top quality, needs A100

General models: Llama 3.1/3.2 (1B–70B), Mistral 7B/Nemo 12B, Gemma 2 (9B/27B).

Pod management

deus setup --runpod    # launch a new pod
deus connect --runpod  # connect to an existing pod
deus setup --stop      # stop pod (stop paying)

Enable auto_stop_runpod: true in ~/.deus/config.yaml to stop the pod automatically after each prompt.

CLI reference

deus "your prompt"              ask anything
deus "prompt" --path ./src      specify working directory
deus "prompt" --no-map          skip repo-map (faster for large repos)
deus "prompt" --model MODEL_ID  override active model

deus setup --runpod             launch a new RunPod pod
deus setup --stop               stop current pod
deus connect --runpod           connect to existing pod

deus model list                 list downloaded + available models
deus model download             download a model to your pod
deus model download --size small  filter by size (small/medium/big/all)

In-chat commands (REPL only):

--model                show active model and downloaded list
--model MODEL_ID       switch active model mid-session

Comparison

	Deus	Claude Code	Cursor
Runs on your GPU	✅	❌	❌
No per-token cost	✅	❌	❌
Code stays private	✅	❌	❌
Any open-source model	✅	❌	❌
Task planning	✅	✅	✅
Web search	✅	✅	✅
Terminal-native	✅	✅	❌
Free to use	✅	❌	❌

Configuration

Config lives at ~/.deus/config.yaml. Created automatically on first run.

base_url: https://your-runpod-endpoint/v1
api_key: your-key
model: Qwen/Qwen2.5-Coder-7B-Instruct
max_tokens: 8192
auto_stop_runpod: false
search_backend: duckduckgo   # or: brave
brave_api_key: ""            # required if search_backend: brave

Requirements

• Python 3.12+
• A RunPod account — or any OpenAI-compatible vLLM endpoint
• GPU with 4 GB+ VRAM (16 GB+ recommended for coding tasks)

License

AGPL-3.0 — free to use, modify, and distribute. Commercial use requires a separate license. See LICENSE for details.