Handwork for macOS and Windows

The coding agentthat does the handwork.

Tell Handwork what to fix or build. It reads your code, makes the change, and runs your tests, with the AI model you choose.

Desktop app v0.1.51

apps in action

CLI tests passed

desktop tests passed

test failures

About this recording

Real Handwork CLI output and desktop app capture, sped up for clarity. In the terminal, Handwork fixes slug.js and passes four tests. In a separate desktop demo project, it creates a slugify function and passes five tests. The silent video loops every 24 seconds.

Works with the model you already pay for

OpenAI
Anthropic
Gemini
xAI
Mistral
DeepSeek
Grok
OpenRouter
Ollama

How it works

What you get

Built to be checked.

Handwork edits files and runs programs on your machine. Every part of it is designed so you can see and control what changes.

Small, reviewable changes

Edits stay focused on the task. Every change shows up as a diff you can read, then commit or throw away.

Reads before it edits

It opens the files that matter and explains the problem before changing anything.

Runs your tests

It runs the commands you ask for and reports what passed and what failed.

You set the permissions

Choose whether Handwork asks before each action, accepts edits on its own, lets a reviewer decide, or works without checks.

Your model, your account

Picks up where you stopped

Conversations are saved, so you can resume a task later with its full context.

Benchmarks

Light on memory.

The runtime is written in Zig. On the same async search task, Handwork peaked at 24.4 MiB of memory. The other agents used 234 to 790 MiB.

0 / 20

Attempts completed across four tasks, rerun October 19, 2026

0.0 s

Median time to fix the async search task

Peak agent memory, async search

Handwork24.4 MiB
Codex234.0 MiB
Claude Code275.2 MiB
OpenCode790.1 MiB

We kept every result, including failures. OpenCode and Codex ran on September 18-19 and Claude Code on October 4, 2026. Handwork, OpenCode, and Codex used GPT-6 Astra at low effort; Claude Code used Opus 5.5 at medium effort. Different models and run dates limit the comparison, and four tasks do not establish an overall ranking.

See every task and metric

Get Handwork

Start with one task.

The desktop app and the command line run the same agent. Install either, connect a model, and describe what you want changed.

Desktop app

v0.1.51

Open a folder, chat with Handwork, approve its actions, then review and commit the changes. The agent is built in.

macOS Apple Silicon
Macs with an M-series chip. Signed and notarized by Apple.
Download
Windows x64
Most Windows PCs with an Intel or AMD processor.
Download
Windows ARM64
Windows on ARM, including Snapdragon laptops.
Download

Release notes

Not sure which Windows build? Check Settings > System > About.

Command line

v0.1.7

Install with npm, open a project folder, and run handwork. Connect a model with /provider.

npm install -g handwork

macOS on Apple Silicon. Windows x64 and ARM64 are experimental.

01
Install the CLI with the command above.
02
Open a project folder and run handwork.
03
Connect a model with /provider, then describe the change.

Read the setup guide

The coding agentThe coding agentthat does the handwork.that does the handwork.

Handwork terminal and desktop demo

You describe the change. Handwork finds the code involved, makes a focused edit, and runs your tests. You review the diff and decide what to keep.