Core Architecture: Autonomous Multi-Agent Loop
Kwabs is engineered around a modular, deterministic multi-agent orchestration loop designed for desktop execution. Unlike single-shot chatbots, Kwabs structures engineering tasks into five distinct phases:
1. Planner Agent
Deconstructs high-level human objectives into discrete, testable milestones with dependency tracking.
2. Coder & Terminal Agent
Writes surgical file diffs and interacts directly with an embedded pseudo-terminal (PTY) session.
3. Researcher Agent
Indexes local codebases, reads configuration files, and performs lexical and semantic searches across project files.
4. Verifier Agent
Executes test suites, linters, and compiler checks to autonomously validate changes before declaring a task finished.
Running Offline Models: Ollama & llama.cpp
Kwabs enables 100% air-gapped, zero-latency local development by integrating directly with Ollama and llama.cpp runtimes.
How Offline Execution Works
- One-Click Engine Lifecycle: Kwabs detects, launches, and monitors local inference daemons natively on your machine without requiring manual Docker or Python environments.
- GGUF Quantization: Runs modern quantized models (Q4_K_M, Q5_K_M, Q8_0) across Llama 3, DeepSeek Coder, Qwen 2.5 Coder, and Mistral.
- Local Embeddings: RAG and memory queries are embedded locally using compact on-device models (e.g. nomic-embed-text or all-minilm).
- Hybrid Fallback: If a developer encounters a high-complexity problem requiring cloud reasoning, they can switch to OpenAI, Anthropic Claude, or Google Gemini by bringing their own API key (BYOK).
The Safety Gateway: Real-Time Command Risk Rating
Autonomous AI agents executing shell commands pose serious operational hazards if left unsupervised. Kwabs implements a multi-layer Safety Gateway that intercepts and evaluates every command before execution.
| Risk Tier | Example Commands | Execution Policy |
|---|---|---|
| Low Risk | ls, git status, cargo check, npm test, cat | Executes automatically with real-time streaming output. |
| Medium Risk | npm install, git push, docker build, pip install | Triggers visual confirmation banner with full parameter inspection. |
| High Risk | rm -rf, sudo, chmod -R, drop database, mkfs | Hard pause. Requires explicit user approval; blocked outside workspace boundaries. |
Additionally, file system write operations are strictly sandboxed to the active workspace directory, preventing accidental modifications to root or parent directories.
The Model Cookbook: Hardware Profiling & Tiers
Rather than forcing developers to guess GGUF quantizations or VRAM bounds, Kwabs includes the Model Cookbook: an automated hardware profiling engine that inspects CPU instruction sets (AVX2, AVX-512, NEON), system RAM, and dedicated GPU VRAM (CUDA, ROCm, Metal) to assign a recommended tier:
8 GB RAM
Integrated graphics or CPU-only execution.
- • Qwen 2.5 Coder 1.5B (Q4_K_M)
- • DeepSeek Coder 1.3B (Q4_K_M)
- • Llama 3.2 1B / 3B
16 GB RAM
6 GB to 8 GB VRAM (RTX 3060/4060, Radeon, Apple M-series).
- • Qwen 2.5 Coder 7B (Q4_K_M)
- • DeepSeek Coder 6.7B (Q4_K_M)
- • Mistral 7B Instruct
32 GB+ RAM
12 GB+ VRAM or unified memory workstations.
- • Qwen 2.5 Coder 14B / 32B
- • DeepSeek Coder 33B
- • Command R+ / Llama 3.1 70B (CPU offload)
Private Long-Term Memory & Security Architecture
Kwabs is built on the philosophy of zero telemetry and zero external data transmission for local tasks. Memory persistence operates strictly within your local environment:
- Embedded Vector Store: Conversation histories, code snippets, and architectural guidelines are tokenized and stored in a local embedded vector database on your disk.
- OS Secure Storage: Third-party API keys (OpenAI, Anthropic, Gemini) and terminal credentials are stored inside your operating system's native secure keyring (Windows Credential Manager or Linux Secret Service / Keyring), with optional biometric authorization.
- No Cloud Training: Your intellectual property, proprietary codebase, and terminal outputs are never transmitted to external servers for training or monitoring.
Installation Guide & Package Formats
Kwabs is distributed as native pre-compiled binaries for desktop operating systems.
Windows
Windows 10 and Windows 11 (64-bit)
- • Native
.exeinstaller with auto-update - • Integrated with PowerShell and WSL2
- • Hardware accelerated via DirectML / CUDA
Linux
glibc 2.31+ on x86-64
- • AppImage: Universal zero-install bundle for all distros
- • Debian / Ubuntu: Native
.debpackage - • Fedora / RHEL: Native
.rpmpackage
Get Started with Kwabs Today
Download the latest build for your system and start coding with local agents in minutes.
Go to Downloads