AIPOCH Open-Science

AIPOCH Open-Science is a desktop research workbench in which scientific AI agents read files, search the web, run Python and R code, query scientific data sources, and produce reports, tables, and figures inside one local project. You create a project, describe a research goal in plain language, attach inputs, pick a model and an approval mode, and review the agent's tool activity, approval requests, and results as they arrive. It runs on macOS, Windows, and Linux, and the same web interface can also run on a standalone Node backend without Electron or a display.

The workbench targets computational and data-intensive research across disciplines such as machine learning, statistics, life sciences, chemistry, materials science, physics, and environmental science, and covers the process from literature review and hypothesis development through code execution, data analysis, simulation, and visualization. Every generated artifact is stored as an immutable, checksummed version whose Provenance view exposes the producing code, execution history, referenced inputs, observed environment, producing conversation branch, and Reviewer findings. Evidence that cannot be verified is marked unavailable rather than inferred.

A completed session exports as a portable .science research package containing selected conversation branches, file versions, Notebook records, and verification evidence. Importing one into another project or computer opens a read-only research replay that executes no code, calls no model, and restores no credentials. A first question creates a writable discussion beside it, and a copy can continue the research with follow-up experiments.

It is model-agnostic at the product level: it connects to built-in cloud providers (OpenAI, Anthropic, DeepSeek, NVIDIA Build, and others), to a custom gateway speaking the Messages, Chat Completions, or Responses protocol (with presets for Ollama, LM Studio, llama.cpp, and vLLM), or to an existing Claude or Codex subscription. The agent backend is Claude Code, OpenCode, Codex, or CodeBuddy, and provider availability depends on which backend is selected. The application is a Node backend with an Electron client built on React, TypeScript, Prisma/SQLite, and an Agent Client Protocol based agent runtime.

Project data, settings, artifact versions, provenance evidence, and API keys stay on the local computer, with keys held in the operating system's credential store where available. Model requests, web searches, and remote connectors still send data to external services, so three permission profiles (ask for approval, auto-approve edits, and full access) bound what an agent may do without a person's confirmation. On the BiomniBench-DA Public 50 comparison it reports a ranking score of 79.05 with gpt-5.6-sol (xhigh), the equal-weight mean of a Gemini 3.1 Pro judge score of 81.04 and a DeepSeek v4-pro judge score of 77.06.