Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/aipoch/open-science/llms.txt

Use this file to discover all available pages before exploring further.

Open Science is a free, open-source, model-agnostic AI workbench for scientific discovery. Built with Electron, React, and TypeScript, it pairs a coordinating AI agent with a persistent Python notebook kernel so researchers can plan multi-step analyses, execute code, store artifacts, and preview results — all from a single desktop application they host themselves, under Apache-2.0.

Science Is Not a Privilege

It shouldn’t require a subscription tier, a supported billing region, or one company’s approval to put AI to work on real research. Peer review doesn’t check your credit card. A hypothesis doesn’t care what currency your lab is funded in. Knowledge has always advanced by being shared, checked, and rebuilt in the open — the tools that now sit at the center of that process are the last thing that should be locked behind a paywall.
That belief is the entire reason this project exists. A university lab in a country without billing access, a hospital that legally cannot send patient data to a third-party API, an independent researcher running everything on a local GPU box, or a team that simply wants to read the code that touches their data — all of them deserve a seat at the table. Open Science is an attempt to build that seat from first principles.

How Open Science Compares

The clearest current articulation of what an AI-native scientific workbench looks like is already on the market — closed source, locked to a single model family, and gated behind a subscription. Open Science is an independent, open implementation of the same problem space. Here’s where each project actually stands:
Claude ScienceOpen Science
SourceClosed sourceOpen source, Apache-2.0
ModelClaude models onlyModel-agnostic — Claude, GPT, Gemini, DeepSeek, Qwen, or a local open-weight model
DeploymentAnthropic-hosted cloudSelf-hosted by default; your infrastructure, your data doesn’t have to leave it
PricingSeat-based subscription (Claude Pro/Max/Team/Enterprise)Free and open; you pay only for the compute/model calls you choose to make
AvailabilityGated by Anthropic billing region and plan tierRuns anywhere you can run the software
Skills~60 curated skills, Anthropic-maintainedOpen skills commons — community-contributed, versioned in git, forkable
Domain scope todayLife sciences (genomics, proteomics, structural biology, cheminformatics)Life sciences, plus social science and economics from day one (planned)
ComputeSSH/HPC access plus Modal for on-demand GPUsPluggable compute fabric — any HPC scheduler, any cloud GPU provider (planned)
Reviewer / verification agentYes, shipping todayYes, planned as an open, inspectable layer
CustomizationConfigure agents inside Anthropic’s product surfaceEvery layer — gateway, skill runtime, compute broker, reviewer — is inspectable and replaceable
MaturityA shipping, polished product, in use todayEarly Alpha — core loop works end to end; scientific differentiators are mostly still ahead
The Maturity row matters: if you need a fully-featured AI research assistant today, Claude Science is the more capable choice. Open Science’s advantage isn’t feature parity yet — it’s the structural ceiling underneath. Nothing about a good AI workbench architecture requires it to be closed, single-vendor, or subscription-gated. Those are business-model choices layered on top of good engineering, and they’re the choices Open Science rejects on principle.

Design Principles

These are the constraints the project will not trade away as it grows. Every design decision is downstream of one belief: science is not a privilege, and the tools built for it shouldn’t behave like one.

Model-Agnostic Core

The orchestrator talks to LLMs through a pluggable gateway. Claude, GPT, Gemini, DeepSeek, Qwen, or a locally-hosted open-weight model behind vLLM/Ollama are all first-class citizens — including routing different agents to different models based on cost and capability.

Local-First & Data-Sovereign

Self-hosting is the default deployment target, not an enterprise upsell. Your data, your compute, your keys — unless you explicitly choose a hosted path.

Reproducibility by Default

Every artifact — figure, table, claim — should ship with the code, environment, and data lineage that produced it. This is meant to be a property of the system, not a discipline researchers must maintain by hand.

Skills Are Plain Files

A skill should be versioned, human-readable, and forkable (markdown + code) — auditable by the researcher who’s trusting it with their analysis, not a binary blob from a marketplace.

Human-in-the-Loop by Construction

New data sources, new compute budgets, and new external credentials require explicit approval. Autonomy is opt-in and scoped, never ambient. Today that’s a single tool-call approval gate; per-scope tiers are planned.

Composability Over Monolith

The target architecture is small, swappable services (model gateway, skill runtime, compute broker, artifact renderer) instead of one inseparable black box, so labs can replace the parts they don’t trust or don’t need.

Trust Is Verified, Not Assumed

A reviewer/verifier agent will eventually check citations, units, and statistical methods before output ships — with its own checks themselves inspectable and open.

Access Is a Right, Not a Privilege

No plan tier, no billing-region allowlist, no corporate approval queue stands between a researcher and the software. If you can run it, you can use all of it. This is the principle every other one on this list exists to protect.

Current Status

Open Science is Early Alpha. The core “plan → execute → produce → preview” loop works end to end today. The properties that would make this a genuinely differentiated, science-grade tool — multi-model routing, provenance, a connector ecosystem, remote compute, a skills commons — are mostly still ahead. The project ships an honest path to the full vision rather than faking a finished product.
What works today:
  • Local Electron desktop app with project and session management
  • Coordinating AI agent runtime via Agent Client Protocol (ACP), with typed activity visualization and a tool-call permission gate
  • Persistent Python notebook kernel with durable run history
  • Artifact file storage organized by project, session, message, and run
  • In-app rendering for CSV, FASTA, HTML, image, JSON, Markdown, and text files
  • Home page and new-project flow
What’s still ahead (planned):
  • Multi-model gateway with per-agent routing (model-agnostic core)
  • Specialist sub-agents for genomics, proteomics, structural biology, cheminformatics, social science, and economics
  • Reviewer/verifier agent for citations, units, and statistical methods
  • Open skills commons, seeded by aipoch/medical-research-skills
  • Data connectors for PubMed, UniProt, PDB, ChEMBL, GEO, arXiv, and more
  • Remote compute fabric (HPC/Slurm, cloud GPU)
  • Verification and provenance layer (lineage graph)
  • CLI/SDK and self-hosted web app
The authoritative, kept-up-to-date breakdown lives in ROADMAP.md in the repository — the capability map there tracks what’s shipping versus still ahead layer by layer.

Get Involved

Open Science is at the stage where architecture decisions are still being made. The best way to have influence is to show up now.
  • 🚀 Ready to run it? Head to the Quickstart to clone, install, and launch in minutes.
  • 💬 Want to talk? Join the Discord — architecture debates, RFC drafts, and skill-writing happen there in real time.
  • 🐛 Found a bug or have a proposal? Open an Issue or Discussion on GitHub.
  • 🐦 Stay updated: Follow @aipoch_ai on X for build-in-public updates and roadmap announcements.

Build docs developers (and LLMs) love