aipoch/open-science

Project: aipoch/open-science Description: Open Science is an open‑source, local‑first, model‑agnostic AI research workbench for scientific discovery. It runs as a self‑hosted desktop app on macOS, Windows, and Linux, letting researchers create projects, describe tasks in plain language, and have an AI agent read files, run Python/R code, search the web, call scientific data connectors, and return reproducible reports, tables, and figures linked to an inspectable activity history. How it works: After installing the app, users create a project and start a session. They describe a goal, attach input data, select a model and approval mode, then send the task. The agent executes the natural‑language request by running commands, Python/R code, editing files, searching, or invoking enabled skills and connectors. Users can approve sensitive actions, inspect the agent’s tool activity, view generated artifacts with provenance, branch conversations to explore alternatives, and continue work across sessions. All project state, artifacts, session history are stored locally; model providers can be built‑in cloud services, custom gateways, or Claude/Codex subscriptions. Key features: Local‑first desktop workspace persistence; natural‑language agent with approvals; multiple model providers (built‑in cloud, custom gateway, Claude/Codex); selectable agent framework; execution Python/R kernels, shell commands, app‑managed environments, remote SSH compute; provenance tracking of versions with evidence; scientific skills (e.g., AlphaFold2, Boltz, Borzoi‑plus skill import from GitHub; built‑in scientific data connectors (PubMed, bioRxiv, ChEMBL, etc.; artifact preview panel (CSV, PDF, images, code, molecular structures, notebook history); project‑scoped command palette; dark mode; open‑source Apache‑2.0 license. Limitations: Requires self‑hosting; not a hosted web service. Model usage may incur costs depending on the chosen provider. Notebook execution is optional and external R package management remains manual. Fully autonomous workflows are limited by required user approvals for sensitive actions. Agent capabilities depend on the selected model and enabled skills/connectors. Generated output does not replace expert judgment and must be validated against primary evidence. Best for: Researchers who want a transparent, locally‑run AI‑assisted environment to run scientific workflows, integrate data and code, and keep fully traceable, reproducible records of their experiments.