Make it easy to automate small, repetitive manual desktop tasks.
Worker Forge is an agent that turns a plain-language task description into a small, single-purpose program that runs on
your own machine. You describe what you want; the Forge interviews you about the details and produces a worker — a
native artifact (.exe on Windows, .app on macOS) that does one job, runs to completion, and exits.
A lot of useful desktop work is small, manual, and repetitive: checking a webpage for a specific change or pulling a daily digest from a set of sources. These tasks stay manual for three reasons:
- Automating them costs more time than they save.
- Most users can't write the script themselves.
- General-purpose tools don't cover them — the tasks are too niche.
Automating them with hosted-LLM solutions has three problems:
- Cost. Hosted-LLM pricing today is subsidized by investor and corporate capital, subject to change, and may become cost prohibitive.
- Availability. Providers shut down or deprecate models on their own schedule.
- Connectivity. Hosted calls require an internet connection at run time while many of these tasks could run on a simple computer in laptop mode.
Example Worker Apps created by this skill.
- Paste the following prompt into your agent
Install skill from https://github.com/hansololz/worker-forge/blob/main/dist/worker-forge.skill - Relaunch the agent.
- Trigger the skill
/worker-forge
Out of scope for v1, on the roadmap:
- Linux support. macOS and Windows are supported today; Linux is next — AppImage/static-binary packaging, a native GTK4/Qt UI, and Secret Service keychain glue so a worker can be forged and run on a Linux box too.
- Workers marketplace. Browse and install workers other people have forged.
- Code-signing. Sign Windows and macOS artifacts so recipients don't see the first-run security warning.
- A remote update channel. Push reforged versions of a worker to recipients without re-emailing the binary.
- Auto-reforge on failure. When a worker stops completing its task — an API changed, a site moved, a model deprecated — Worker Forge reforges it from the original spec until it works again.
- A desktop UI for the Forge itself. Today the interview happens in a chat; a desktop app would open the same flow to people who don't use a chat client.
- A CLI surface for workers. Skip the interview and pass the spec on the command line.
- Automated security scanning. The Forge produces source you can read, but a scanner would catch the obvious classes of mistake before the build.
- Cross-platform scheduling helper. Native schedulers — Windows Task Scheduler, launchd, cron — all work differently. A thin cross-platform scheduler shipped with each worker could remove that step.
- Artifact attestation. A built binary should be verifiable against the source in the Workspace, so a recipient knows what they're running.
- Smarter local-model selection. The Forge already proposes the currently-popular model, picks the runtime to match it (Ollama, Hugging Face, …), and can let a GUI worker choose the model at run time. The remaining step is fully automatic selection — the runtime querying the host for installed models and hardware and picking the best available on its own, rather than the user deciding.
design.md— what a worker is and isn't, the cascade as runtime contract, the Workspace layout, the forge / run / reforge lifecycles, key design decisions, and known failure modes.spec.md— the worker-forge skill spec: what the skill must do in each phase (interview, cascade design, code generation, packaging) and how it should behave. Referencesdesign.mdfor the rationale.
MIT. See LICENSE.