OpenScience — Self-Host Your Own AI Research Laboratory (Complete Guide)
OpenScience — Self-Host Your Own AI Research Laboratory (Complete Guide)
Scientific research involves a predictable loop: read the literature, formulate a hypothesis, design an experiment, run it, analyze the results, and write it up. Each step is labor-intensive, and running the full loop for a single question can take weeks.
OpenScience, from Synthetic Sciences, an AI research lab based in San Francisco, collapses this loop into a continuous AI-driven workflow. You give it a research goal, and it reads the papers, writes and runs code, executes experiments on real compute, queries scientific databases, and produces a written summary — all on your own machine using your own API keys.
Launched on July 3, 2026, OpenScience has already gathered over 1,500 GitHub stars and is the fastest-growing open-source AI research tool of the month. It supports any frontier or open-weight model from Anthropic, OpenAI, Google, and 75+ other providers, and comes with 290+ built-in skills covering ML training, biology, chemistry, literature review, and LaTeX.
GitHub: synthetic-sciences/OpenScience | Website: openscience.sh | Docs: openscience.sh/docs
Why OpenScience Matters
The gap between asking a research question and getting an answer is enormous. A typical workflow involves:
- Searching PubMed and arXiv for relevant papers
- Reading and synthesizing findings
- Writing code to test a hypothesis
- Running experiments (possibly on cloud compute)
- Plotting results and drawing conclusions
- Writing up the findings in a paper-ready format
Each step requires switching between tools — browsers, terminals, Python notebooks, LaTeX editors, reference managers. OpenScience replaces this fragmented workflow with a single browser workspace that handles the entire loop.
The key insight is that LLMs are good at research planning, but terrible at tool coordination. OpenScience bridges this gap by giving the AI agent a proper toolbelt: a shell, an editor, LSP integration, MCP servers, 30+ scientific database connectors, and a skill system with 290+ pre-built research capabilities.
Architecture Overview
The system is a local-first architecture running entirely on your machine:
- CLI / Launcher —
opensciencestarts a local Hono server and opens a SolidJS workspace in your browser. The first run walks you through key setup (API keys, optional Atlas account). - Agent Runtime — The core message loop with model routing. Each request goes directly to your chosen provider (Anthropic, OpenAI, Google, or 75+ others). No data leaves your machine unless you configure it.
- Tool Layer — Four categories of tools: shell (execute commands), editor (modify files), LSP bridge (code intelligence), and MCP client (external tool integration).
- Science Connectors — 30+ queryable databases: UniProt, PDB, Ensembl, ChEMBL, PubChem, arXiv, OpenAlex, Semantic Scholar, and more. The agent queries these natively.
- Skills Engine — 290+ instruction bundles loaded on demand. Covers ML training (DeepSpeed, PEFT, TRL), evaluation, dataset work, molecular biology, clinical biology, cheminformatics, papers and LaTeX, figures, and cloud compute (Modal, Tinker).
- Workspace UI — SolidJS frontend with a file tree, editor, terminal, session history, and inline rendering for molecules, protein structures, genomes, and plots.
- Atlas (Optional) — Synthetic Sciences' managed platform for curated frontier models billed from a prepaid wallet. Bring-your-own-key usage is always free.
The server binds to 127.0.0.1 only and enforces a Host/Origin allowlist. There is no remote mode — everything stays local.
Prerequisites
- Node.js or Bun — Bun 1.3+ is recommended. The CLI is a Bun/TypeScript application compiled to a native binary.
- An API key — Any provider: Anthropic (
ANTHROPIC_API_KEY), OpenAI (OPENAI_API_KEY), Google (GEMINI_API_KEY), or 75+ others - 8 GB RAM minimum (16 GB+ recommended for complex research sessions)
- macOS, Linux, or Windows — Platform binaries are attached to GitHub Releases
Installation
OpenScience installs via npm as a single global package:
npm install -g @synsci/openscienceThen start the workspace:
openscienceThe first time you run it, a short setup walks you through model configuration. You can set an API key before starting:
export ANTHROPIC_API_KEY=sk-ant-...
openscienceIf you prefer not to install globally, use the one-shot launcher:
npx synsciThis works identically — it installs and runs OpenScience in a single step. Platform binaries are also available on the GitHub Releases page.
To point the workspace at an existing project:
openscience ~/code/my-research-projectConfiguration
Global configuration lives at ~/.config/openscience/openscience.json. Project-level configuration goes in openscience.json or a .openscience/ directory at the repo root.
Key configuration options:
- Provider keys — Set via environment variables (
ANTHROPIC_API_KEY,OPENAI_API_KEY, etc.) or through the workspace UI's Credentials panel - Model selection — Switch between providers per-request from the model selector in the workspace
- Custom agents — Define custom agent personas in the project config
- Plugins — Load MCP servers and custom tools via the plugin system
- Skills — Install additional skill packs from the Atlas skill index
Using a Local Model
You can use any OpenAI-compatible local endpoint (Ollama, vLLM, LM Studio) by setting the provider URL:
export OPENAI_API_KEY=not-needed
export OPENAI_BASE_URL=http://localhost:11434/v1
openscienceOpenScience routes model requests per-invocation, so you can mix providers freely — use Claude for literature review and a local model for code generation, for example.
Quickstart: Run Your First Research Session
1. Start the Workspace
openscienceThis opens http://127.0.0.1:3000 in your browser.
2. Pick a Model
Use the model selector in the workspace header. You can choose any provider key you've configured.
3. Give a Research Goal
In the chat panel, enter something like:
"Compare the binding affinity of approved EGFR inhibitors and suggest which one has the most favorable selectivity profile. Use ChEMBL and PDB."
OpenScience will:
- Search ChEMBL for binding data on approved EGFR inhibitors
- Query PDB for available crystal structures
- Write and run analysis code
- Generate comparison plots
- Summarize the findings in a structured report
4. Use Specialized Agents
Switch between agent modes depending on your task:
| Agent Mode | Use Case |
|---|---|
research |
Default — full research loop |
biology |
Biological data analysis |
physics |
Physics simulations and analysis |
ml |
ML training and evaluation |
plan |
Read-only planning — generate a research plan without executing |
Select the agent from the workspace UI or configure it in your project's openscience.json.
Working with Scientific Databases
OpenScience exposes 30+ scientific databases as native tools. The agent queries them automatically when relevant:
- UniProt — Protein sequence and functional information
- PDB — Protein Data Bank for 3D structures
- Ensembl — Genome browser data
- ChEMBL — Bioactive drug-like molecules
- PubChem — Chemical compounds and assays
- arXiv — Preprint server
- OpenAlex — Scholarly citation data
- Semantic Scholar — Academic paper search
Example query the agent can handle:
"Find all human kinases with known inhibitors in ChEMBL, retrieve their UniProt entries, and check if any PDB structures exist for the kinase domains."
The agent chains these queries automatically without you writing a single API call.
Skills System
OpenScience ships with 290+ skills organized into domains. Skills are instruction bundles that teach the agent how to perform specific tasks:
- ML Training — DeepSpeed, PEFT, TRL integration for model fine-tuning
- Evaluation — Standard ML evaluation benchmarks
- Dataset Work — Data loading, preprocessing, and augmentation
- Molecular Biology — Sequence analysis, BLAST, alignment tools
- Clinical Biology — Clinical trial data analysis
- Cheminformatics — Molecule manipulation, SMILES parsing, docking
- Papers & LaTeX — Paper writing, reference management, BibTeX
- Figures — Publication-quality plotting and visualization
- Cloud Compute — Modal and Tinker integration for remote execution
Users can install additional skill packs and create custom skills using the TypeScript SDK.
Verification Checklist
After following the setup steps, verify everything works:
opensciencestarts the workspace athttp://127.0.0.1:3000- The browser workspace loads with the chat panel visible
- Model selector shows your configured provider
- Setting a research goal produces a plan, not an error
- Scientific database queries return results (e.g., "Search ChEMBL for aspirin")
- Code execution works (the agent can write and run Python/Shell commands)
- Switching between agents changes the agent's behavior
- Session history persists across restarts
Comparison: OpenScience vs. Alternatives
| Feature | OpenScience | Claude Science | ChatGPT Research |
|---|---|---|---|
| License | Apache 2.0 (open-source) | Proprietary | Proprietary |
| Self-hosted | ✅ Yes | ❌ No | ❌ No |
| Local models | ✅ Any OpenAI-compatible endpoint | ❌ No | ❌ No |
| Data privacy | ✅ Everything runs locally | ❌ Data sent to Anthropic | ❌ Data sent to OpenAI |
| Research agents | 5 specialized agents + custom | 1 default agent | 1 default agent |
| Scientific DBs | 30+ native connectors | Limited | Limited |
| Skills | 290+ built-in, extensible | No | No |
| Plugin system | ✅ MCP + TypeScript SDK | ❌ | ❌ |
| Cloud compute | ✅ Modal, Tinker | ❌ | ❌ |
| Cost | Free (BYO API keys) | Subscription | Subscription |
Resources
- GitHub: synthetic-sciences/OpenScience
- Website: openscience.sh
- Documentation: openscience.sh/docs
- Atlas Platform: app.syntheticsciences.ai
- NPM Package: @synsci/openscience
- License: Apache 2.0
- Built with: Bun, TypeScript, Hono, SolidJS