Open-Science

Open-Science is an open-source, local-first, model-agnostic AI research workbench developed by AIPOCH for scientific discovery. It brings together coordinator and specialist AI agents, Python and R code execution, scientific data connectors, reviewer checks, and traceable research artifacts into a single desktop workspace available on macOS, Windows, and Linux. The software is distributed under the Apache License 2.0 and does not require a seat license, though external AI model providers, scientific services, or compute resources may carry separate costs.

The workbench is designed to keep research questions, project files, code execution, outputs, and review evidence together inside persistent projects. Agents can continue multi-step investigations using specialized capabilities without separating the reasoning interface from the execution environment, supporting a workflow that moves from defining a research question, through planning and delegation, to execution and final inspection of outputs alongside provenance records.

Execution and Compute

• Provides persistent Python and R kernels that allow researchers and agents to maintain variables and analytical state across related steps.
• Supports stateless shell execution with recorded history for file operations, scientific tooling, and reproducible workflows.
• Exposes environment and package information to make computational assumptions more visible.
• Supports SSH workflows for submitting and managing research jobs on registered remote compute hosts when local resources are insufficient.

Specialists, Skills, and Scientific Connectors

• Includes 24 built-in Scientific Connectors covering scientific literature, biomedical databases, genomics, chemistry, clinical research, and related resources.
• Connector and tool access is governed by workspace permissions; researchers can also add compatible custom connectors.
• Specialists are purpose-scoped AI agent profiles configured with selected instructions, Skills, Scientific Connectors, and permissions. The main agent can delegate scoped work to a Specialist, but Specialists cannot bypass the researcher's permission settings.
• Reusable Skills are available for literature review, computational biology, biomolecular modeling and design, environment and package workflows, and remote compute.
• Researchers can use personal and imported Skills through the Skills system in addition to featured built-in Skills.

Traceability and Review

• Research outputs — including reports, tables, figures, and notebooks — are preserved as versioned research artifacts connected to the evidence behind their creation, rather than treated as isolated chat outputs.
• Provenance records connect an artifact to available inputs, producer code, execution history, environment details, and the conversation branch that produced it.
• Reviewer checks compare completed agent work with available transcripts, execution records, and artifacts to identify unsupported claims, inconsistencies, or missing evidence.
• Explicit evidence gaps are surfaced when provenance information is unavailable, rather than reconstructing or guessing what happened.
• Reviewer findings and provenance records improve transparency and auditability but do not certify scientific correctness or replace expert validation.

Data Handling and Local-First Design

• Project state, sessions, uploads, notebook history, and generated artifacts are stored on the user's computer by default.
• Data may leave the device when a researcher invokes a configured model provider, Scientific Connector, web search, or remote compute host; these external calls are governed by active approval and permission settings.
• Provider-specific retention and training policies may still apply, and users are advised to review permissions and runtimes before using sensitive or regulated data.

Model Support and Integrations

• Supports multiple model providers and agent backends, including OpenAI (API and Codex subscription), Anthropic (API and Claude subscription), xAI (Grok) via API or OAuth, and DeepSeek via API.
• Also supports DeepSeek, Bailian, Zhipu AI (GLM), Kimi (Moonshot), MiniMax, StepFun, Xiaomi MIMO, SenseNova, Volcengine Ark, Tencent, NVIDIA, OpenCode, OpenRouter, and Apodex.
• Compatible custom gateways and self-hosted model deployments are supported through a compatible API option.
• Available models depend on the user's account, selected agent backend, API protocol, region, and installed version of Open-Science.

Open-Science is positioned as a research support tool rather than a replacement for scientific judgment. Researchers remain responsible for methods, data governance, interpretation, validation, and final decisions. The latest release, including installers, compatibility information, release notes, and known limitations, is available on the Open-Science GitHub release page.