An open-source workbench for answering research questions and tracking how results were produced
When you open a paper or data file and ask how a specific result was derived, an AI agent executes code, generates charts, and displays the process transparently. AIPOCH Open-Science is an open-source desktop application for scientific research that handles data locally and allows you to connect any model. It records the inputs and code used to create each output, ensuring reliability when you review results later or share them with colleagues.
v1.0.8 Update and Recent Changes
The latest version, v1.0.8, was released as of November 2026. This update reorganized the settings screen into tabs by function for easier navigation. It added the ability to import up to 100 documents at once using DOI, arXiv, and PubMed identifiers, and enabled viewing AI-generated research snapshots in the document detail view. Additionally, agent processes now release properly upon session completion, and the update process on macOS and Windows has been improved for smoother operation.
Transparency and Reproducibility of the Research Process
The core of this tool is revealing the process behind the results. Charts, reports, and tables created by the AI are all stored as immutable versions, with each version recording the code used for generation, execution history, input data, and environment information. You can inspect this evidence directly through the 'Provenance' view, and any unverifiable parts are marked as 'unavailable' rather than guessed. Recorded research sessions can be exported as .science packages to be imported into other computers or projects. Imported records remain read-only, and code execution or credential restoration does not occur, ensuring safety.
Extensible Scientific Tools and Model Selection
You can extend research workflows using 24 built-in scientific skills and 525 skills from the marketplace. Access scientific data sources via 48 built-in connectors and custom MCP connectors, with fine-grained permission management for each tool. Model selection is also flexible, supporting cloud LLM providers, custom gateways, or existing Claude and Codex subscription accounts. Python and R notebooks can be executed locally, or jobs can be submitted to remote HPC via SSH or Slurm.
Performance Evaluation and Benchmark Results
The README publishes self-evaluation results from various benchmarks, including AstaBench, ASI-Bench, and BiomniBench-DA. In the AstaBench LitQA2-FullText-Search test, the gpt-6-astra model achieved a Recall@30 of 90.67%, tying the highest score reported at the time. In the ASI-Bench B1 through B4 comparison, the GPT-6 model scored 79.35, 56.11, 59.67, and 59.46 respectively, recording the highest scores in that comparison group. Since tasks and grading methods differ across benchmarks, direct comparison between results is difficult.
Installation Environment and Pre-Use Checks
The software supports macOS 12 or later (Apple Silicon and Intel), Windows 10/11 x64, and Linux x64/ARM64. On macOS, it can be installed via Homebrew, while the standalone CLI requires Node.js 22.13.0 or higher. Research data is preserved during Windows reinstallation, but a data reset tool is required for a complete initialization. This tool has a local-first structure, making it suitable for researchers who prefer environments where data is not sent to external servers. However, large files may require split processing depending on model context limits or file size (up to 10 GiB), and you should verify host and software permission settings before using remote computing.
