An AI workspace that actually gets work done.
Kwabs plans, codes and runs commands in a real terminal, asking before anything risky. Run models locally with Ollama or llama.cpp, or bring your own OpenAI, Anthropic or Gemini key. Your memory stays on your machine.
Available for Windows and Linux
Everything an assistant needs to do the work, not just talk about it.
Autonomous Agents
An orchestrator hands work to specialised planner, coder, researcher, terminal and verifier agents, then tracks the plan step by step. Pause or stop at any point.
On-Demand Skills
Teach agents specialized workflows from an open library of playbooks. Loaded dynamically when relevant to avoid prompt bloat, and stored globally in ~/.kwabs/skills.
Safety Gateway
Every terminal command is rated low, medium or high risk. Risky ones wait for your approval, sudo prompts are answered in-app, and file tools are sandboxed.
Local or Cloud Models
Run models fully offline with Ollama or llama.cpp, installed in one click. Or switch to OpenAI, Anthropic or Google Gemini whenever a task calls for it.
Semantic File Search
Search documents, PDFs, photos and recordings by meaning with /search. Local embeddings index your folders in the background with zero data leakage.
Model Cookbook
Kwabs scans your CPU, RAM and GPU, places your machine in a performance tier and recommends the models that will actually run well on it.
Live Canvas
Watch agent output render beside the chat: Markdown docs, live HTML previews, SVG artwork and Mermaid diagrams, ready to inspect and copy.
Private Long-term Memory
Local embeddings and on-device vector search let Kwabs recall past conversations. The index never leaves your machine, even when you chat with cloud models.
PC Doctor Diagnostics
Built-in cross-platform system diagnostics for real-time CPU/GPU load, memory usage, disk health, process triage, and automated OS troubleshooting.

Built for professionals. Designed for everyone.
Everything you need to master Kwabs.
Getting Started
Download the installer for your platform and launch Kwabs. Open a project folder to give the agents a workspace, then describe what you want done in the chat.
Choosing a Model
Open Settings to add API keys for OpenAI, Anthropic or Gemini, or install Ollama or llama.cpp in one click. Not sure what your machine can run? The Model Cookbook recommends models for your hardware.
Working with Agents
Kwabs breaks bigger requests into a plan and works through it in a built-in terminal. You can queue follow-up messages, pause between steps, or press Esc to stop the agent.
Agent Skills & Playbooks
Equip agents with expert playbooks for coding, data analysis, finance, and system operations. Built-in and community skills from skills.sh load dynamically on demand.
Local File Search
Search files, PDFs, images, and audio by meaning with /search and jump straight to matches with /goto. All processed locally with background folder watching.
Privacy & Security
With local models nothing leaves your machine. Conversation memory is always indexed locally, API keys are stored only on your device, and saved terminal passwords are encrypted with OS secure storage.
Keyboard Shortcuts
- Ctrl + N
- New conversation
- Ctrl + K
- Focus message box
- Ctrl + H
- Conversation history
- Ctrl + `
- Toggle terminal
- Ctrl + ,
- Open settings
- Esc
- Close panel / stop agent
Frequently Asked Questions.
Everything you need to know about Kwabs, local models, privacy, and system requirements.
Evolving day by day.
v1.5.2
Latest- Zero-friction on-demand agent skills: all built-in and installed domain playbooks are enabled by default and dynamically matched without prompt bloat.
- Global skills storage: installed skills live universally in ~/.kwabs/skills across all workspaces and projects.
- Open agent skills ecosystem integration: discover and install community playbooks directly from skills.sh via the built-in find-skills tool (npx -y skills add <package> -g -y).
- PC Doctor diagnostic skill: real-time cross-platform hardware diagnostics, OS performance checks, resource monitoring, and automated troubleshooting.
- Fine-grained token usage tracking across all model providers and customizable agent output conciseness styles.
v1.5.0
- Semantic local file search: search documents, PDFs, photos, and audio recordings by meaning using local embeddings without uploading anything.
- Instant navigation with /goto: jump directly to files, folders, line numbers, or URLs with integrated side-panel preview.
- Background folder indexing: automatic file watching with incremental updates and visual status tracking.
- Voice control for Gemini Live sessions: hands-free voice commands to start, pause, and navigate active sessions.
v1.4
- Rewritten in Rust with Tauri: the Windows, .deb and .rpm installers are now 5–7 MB instead of 150–360 MB, and the app uses about a fifth less memory.
- Native model clients for OpenAI, Anthropic, Gemini and Ollama, with streaming and tool calling.
- llama.cpp runs as a managed llama-server, downloaded on demand instead of bundled.
- Memory search runs on-device in Rust; existing conversations and memories carry over.
- Faster built-in file search that respects .gitignore.
v1.2
- Multi-agent architecture: an orchestrator routing work to planner, coder, research, terminal and verifier agents.
- Pause agent loops between steps, and queue messages while one is running.
- Live voice sessions with Gemini Live and OpenAI Realtime, including screen sharing and a floating mini-dock.
- Offline speech-to-text powered by whisper.cpp.
- Encrypted credential storage with biometric unlock for terminal password prompts.
- Long-term memory moved to sqlite-vec for faster local vector search.
- Skill catalog and installer for extending what agents can do.
- Light theme, file previews, Fedora RPM packages and cross-built Windows installers.
v1.0
- Initial public release of Kwabs.
- Native support for OpenAI, Anthropic and Google Gemini.
- Local model integration with Ollama.
- Vector-based long-term memory.
Report a bug or share an idea.
Messages go straight to the developer. Leave your email if you'd like a reply.