Open-source desktop app for social scientists: turn a published human study into a reviewable, replicable Human Program with an agent, and check every claim against the original paper.
- Three panes, each foldable: the study agent, the original paper, and the Human Program (a readable protocol plus generated artifacts).
- Projects live as plain folders in
~/Documents/HumanStudy Projects. - Studies: a community library backed by HumanStudy-Bench; fork a study into a local project.
- Kernel-agnostic: the agent runs on the HumanStudy edition of opencode today; the
social-science layer in
pack/(AGENTS.md and skills) works with any AGENTS.md-aware harness.
Status: early development. See SPEC.md.
cd app
npm install
npm run dev # Electron
npm run preview:web # renderer in a browser with a scripted agent
npm run typecheck
npm run dist:mac # signed macOS build (needs a Developer ID certificate)Nothing to install: on first run the app downloads its agent kernel, the HumanStudy edition of
opencode, from the HumanStudy-Hub/opencode releases
(pinned to one tag, verified against the release's SHA256SUMS.txt) into its own data folder. The
app never uses an opencode installed on your machine, and the kernel keeps its config, sessions and
caches under the app's data folder (kernel/), so a personal opencode setup is untouched. For
development, HUMANSTUDY_KERNEL_BIN=<path> runs a locally built kernel and HUMANSTUDY_KERNEL=mock
uses the built-in mock agent.
The dev branch of the website deploys to a protected Vercel preview. npm run dev:dogfood
points the app at it and sends the deployment-protection bypass secret, which it reads from the
macOS Keychain item humanstudy-dev-bypass (ask a maintainer for it; it is never committed). Use
the dev promo code to unlock hosted models there.