A chat is a workflow
Every chat runs one selected workflow from start to finish — branching, looping, calling tools and delegating to sub-agents while you simply watch or step in.
v0.1.0 — early access
Apache-2.0 Open Source
Windows + Linux
The open-source AI agent desktop app.
Like Claude Code or Codex — an AI agent that works right on your computer with real tools for files, shell and the web. Except Aworkit is free and open source, and every chat is a visible workflow you can open, inspect and reshape with a mouse — using any model, local or hosted.
Runs entirely on your machine — no cloud, no account. klutzGames main site · Legal notice
158 model calls · 246 tools · 99.2% cache hit
What it does
A familiar chat app — with the engine exposed.
Aworkit combines the comfort of a desktop chat window with a graphical editor for complete agent harnesses. Technically capable, deliberately simple to use.
Visual workflow editor
Drag nodes onto a canvas and connect them — agents, conditions, loops, approvals. Workflows are plain, portable JSON you can export, diff and share.
Your models, your choice
Local model, hosted API or external coding agents like Codex — all are first-class citizens. Run fully offline, fully hosted, or any mix in between.
Tools, MCP & external agents
Built-in file, shell, Python and web tools — plus any MCP server and lifecycle-owning agents such as Codex or Claude Code, configured right in the app.
Approvals & transparency
Actions that write files, run commands or spend money can require your explicit approval. Every model call, tool action and error is recorded and inspectable.
Local-first & private
Aworkit is a desktop application, not a web service. Chats, workflows and settings stay on your machine; credentials go into the operating system's keychain.
How it works
From idea to running harness in three steps.
Pick a workflow
Start with a simple chat or a ready-made template, or open the editor and build your own graph of steps — no code required.
Chat & watch it run
Type your request. The workflow resolves its models and tools and executes step by step — you can pause, steer, approve or cancel at any time.
Inspect everything
Open the Run details panel to see token usage, cache hit rates, every tool call and the raw JSON behind each step. Nothing stays hidden.
Screenshots
Look inside.
Click any screenshot to view it full size.
Get started
Two ways in.
Installer
Grab a prebuilt package from the GitHub releases page and install it like any desktop app.
- Windows .msi installer or setup .exe
- Linux .deb, .rpm or portable .AppImage
Build from source
Requires Rust 1.97+, Node.js 20+ and pnpm. Full instructions for Windows and Linux are in the README.
# build the core, then the desktop app
cargo build --workspace --release --locked
cd desktop
pnpm install --frozen-lockfile
pnpm desktop:build ✓
Read the README ↗
Open source
Free as in freedom.
Aworkit is developed in the open and built on well-known open-source libraries.
APACHE-2.0
Use it, modify it, ship it — commercially or not. The full source, issue tracker and contribution guide live on GitHub. Contributions are welcome.
github.com/fuleshu/Aworkit ↗Developed with Adashi
This project is built using Adashi — the senior agent dashboard — which provides the design model, tasks and QA workflow that AI coding agents follow in this codebase. Adashi can also plug into Aworkit as an MCP server.
Adashi on GitHub ↗Ready to look behind the curtain?
Aworkit is free, open source and runs on your own machine. Download it, or browse the source first — whatever suits you.