OpenScience
Synthetic Sciences · OpenScience: Open-Source AI Research Agent and Workbench for Scientific Computing
OpenScience is an open-source (Apache 2.0) research agent and workbench from Synthetic Sciences. Given a goal, it reads the literature, writes and runs analysis code in shell, Python, and R kernels, runs experiments locally or on remote compute, and writes up results, with every step recorded. It suits scientists, ML researchers, and computational labs that want an agent working on their own files with their choice of model.
Best for
Computational scientists, ML researchers, and research engineers who want an open-source agent that runs analysis code and experiments on their own files and compute, with a full record of each step and their choice of model provider
Not ideal for
Researchers who only need to search and summarize papers, people who want a hosted service with no account or local install, and teams that cannot accept a young, fast-changing tool or default trace sharing without changing settings
Who it's for
Scientists, ML researchers, bioinformaticians, chemists, and research engineers working in academic labs, startups, and R&D teams
Capabilities
- Research agent that plans, searches the literature, writes and runs code and experiments, and writes up results, with a visible trace of every search, command, and file it produced
- Shell, Python, and R kernels, notebooks, and a project file system with explicit read and write grants for anything outside the project
- Search across registered scientific databases such as UniProt, ChEMBL, PubMed, and arXiv, with fetches of FASTA, PDB, CIF, or SDF files where a source supports them
- Hundreds of bundled skills across biology, chemistry, physics, ML, and data engineering, with attribution to the open collections they come from
- Delegation to parallel workers: a read-only explore scout, ML, biology, physics, chemistry, and data specialists, and a general worker
- Remote compute on Modal (GPU types from T4 to B200, up to 24 hours per job), plain SSH hosts, or Slurm and PBS clusters, with approval prompts for remote and paid jobs
- Autoresearch studies that hill-climb a metric over many runs with a baseline, ranked ideas, kill criteria, a budget, and readable study, ideas, results, and lessons files
- Experimental adapters for ten NVIDIA BioNeMo NIM models, including Boltz-2, DiffDock, Evo 2, OpenFold3, ProteinMPNN, and RFdiffusion, using your own NVIDIA API key
- Normal and Ultra research effort, a plan mode that only writes its plan files, and permission levels from always-ask to full access, with network access approved per host
- Model access through your own API keys, a supported ChatGPT sign-in, local models (Ollama, LM Studio, or any compatible endpoint), or Ace pay-as-you-go managed models
- Extensible with remote and local MCP servers, custom agents and commands, plugins, and a TypeScript SDK; `openscience run` supports scripted, single-turn use
Limitations
- First-run setup requires a Synthetic Sciences account, even when you use your own provider keys or a local model (headless `openscience run` does not need it)
- While signed in, session traces, which can include prompts, tool inputs and outputs, and searches, are shared by default, including sessions on your own keys or local models, until you turn sharing off
- Skills are instructions, not installed tools: many workflows need additional software, data, or accounts before a substantial run
- A local model keeps only inference on your device; web search, online databases, remote compute, and trace sharing can still send data off the machine
- Model, compute, and external service costs are separate from the free workbench, and long conversations, extra research branches, and Ultra effort increase usage
- Database searches return at most 50 hits per request, so a single search is not a full database export
- Published benchmark scores were run by the project itself with GPT-6 Astra and GPT-6 Sol, although the traces are public
- A young project: the repository was created in July 2026 and ships releases almost daily, so behavior and settings change often
Use cases
- Checking a CSV dataset for missing values and inconsistent labels, saving a quality report, a plot, and the code to reproduce them
- Reproducing a paper's claimed result as a bounded experiment and recording where the measured result differs
- Running an overnight hyperparameter or architecture study on a training script with Autoresearch and a fixed budget
- Writing a literature review with a saved search record, an evidence table, and cited sources
- Looking up a protein in UniProt, fetching its sequence, and running structure or binder-design steps through BioNeMo endpoints
- Sending a GPU training job to Modal or a Slurm cluster and pulling the outputs back into the project
Our take
OpenScience focuses on the computational half of research, not just the reading. The combination of real kernels, scientific databases, cluster and Modal dispatch, and an Autoresearch loop with a written ledger is useful for anyone who reproduces results or runs many experiments, and the trace of every step makes the output auditable. Two settings deserve attention before serious use: it needs a Synthetic Sciences account to start, and trace sharing is on by default while signed in, which matters for unpublished or sensitive data. Expect frequent changes, since the project is only a few months old.
Who should use it
Computational biologists, chemists, physicists, and ML researchers who want an agent to clean data, run analyses, reproduce results, and manage experiment loops on their own machines or clusters, and labs that want an open-source tool they can inspect and extend.
Who should skip it
Researchers who only need paper search and summaries, teams that need a hosted tool with no local install, and groups with strict data rules that cannot change default trace sharing or accept a fast-moving early project.
Strengths
- Free and open source under Apache 2.0, running on your own machine and files
- Executes real analysis code and experiments, not just literature summaries
- Every step is recorded, so methods and results can be checked
- Works with your own keys, a ChatGPT sign-in, local models, or pay-as-you-go Ace
- Remote compute on Modal, SSH hosts, and Slurm or PBS clusters, with approvals
Weaknesses
- Requires a Synthetic Sciences account for first-run setup
- Session traces are shared by default while signed in
- Many skills need extra software or accounts to actually work
- Very new and changing quickly
OpenScience pricing
Workbench
$0
- Free and open source (Apache 2.0)
- Desktop app, CLI, and browser workspace
- Synthetic Sciences account required for first-run setup
Your own provider or local model
Provider rates
- Your API key or supported sign-in is billed by that provider
- Local models run on your hardware with no Ace model charge
Ace
Pay as you go
- No monthly subscription; enabling Ace costs $0
- Per-model token rates shown in the app before you start
- Optional auto reload of a fixed $20 when purchased funds fall below $5
Free tier limits: The workbench is free with no usage limits of its own; model, search, and compute costs depend on how you connect them.
Note: Ace charges the serving provider's token rates. Direct provider routes add no fee; models served through OpenRouter include its funding fee (5.5% by default) in the displayed rate, and card processing fees are shown at checkout. Managed research search also draws on the Wallet. Monthly usage limits and reload limits can be set in billing. Compute providers such as Modal bill separately under their own terms.
Technical specs
API pricing
Workbench free; Ace is pay as you go at per-model token rates shown in the app, and your own keys are billed by your provider
Available models
Where OpenScience excels
Reproducing a published result before building on it
OpenScience agrees the claim, prerequisites, and budget first, then runs the experiment and records how the measured result compares, so a lab knows whether the baseline holds before investing more time.
Running an overnight model-tuning study
Autoresearch keeps a baseline, queues ideas, ends runs that break kill criteria, and writes a ledger of kept and reverted changes, which turns a long sweep into a reviewable record.
Auditable data analysis for a methods section
Each turn saves the code, figures, and trace behind a result, so the analysis can be rerun and the methods described accurately.
How OpenScience fits alongside other tools
OpenScience and Elicit
Elicit searches and screens scientific literature, extracts data from papers into tables, and supports systematic reviews. OpenScience can review literature too, but its focus is running the analysis, code, and experiments on your own data. A lab can screen evidence in Elicit and then use OpenScience to reproduce or extend the analysis.
OpenScience and Claude Code
Claude Code is a general coding agent for software repositories. OpenScience adds research-specific pieces around a similar agent loop: Python and R kernels, scientific databases, bundled domain skills, cluster and Modal job dispatch, and Autoresearch studies. Teams can keep Claude Code for building research software and use OpenScience for analysis and experiment work.
Frequently asked questions
What is OpenScience?
OpenScience is an open-source AI research agent and workbench from Synthetic Sciences. You describe a research task in plain language, and it plans the work, searches the literature and scientific databases, writes and runs code and experiments, and hands back results with a record of every step. It runs as a desktop app, a CLI with a browser workspace, or headless.
Is OpenScience free?
The workbench is free and open source under Apache 2.0. Models, compute, and external services can cost money. You can use your own provider API keys (billed by your provider), a local model (no model charge), or Ace, which is pay as you go from a Wallet with no monthly subscription. Optional automatic reloads add a fixed $20 when purchased funds fall below $5.
Can OpenScience run with local models?
Yes. You can connect Ollama, LM Studio, or another compatible endpoint, and local model requests are not charged to Ace. Only inference stays local: web search, online databases, remote compute, and signed-in trace sharing can still send data off your device.
Do I need an account to use OpenScience?
Yes for first-run setup, which asks you to sign in to a Synthetic Sciences account even if you use your own keys or a local model. Headless `openscience run` does not require sign-in.
Does OpenScience share my data?
While you are signed in, session traces are shared by default, including sessions that use your own keys, subscriptions, or local models. Traces can include prompts, model responses, tool inputs and outputs, and searches. You can turn sharing off under Customize, General, Data and privacy, or disable it at the account level.
Can OpenScience use GPUs or a cluster?
Yes. It can dispatch approved jobs to your own Modal account (GPU types from T4 to B200, up to 24 hours per job), to plain SSH hosts, or to Slurm and PBS clusters, then follow their status and retrieve outputs into the project.
Integrations & fit
About OpenScience
OpenScience is built for the parts of research that are real work but not the idea: pulling and cleaning data, reproducing a published claim, sweeping a parameter, checking references, and drafting a methods section. It runs as a desktop app for macOS, Windows, and Linux, as a CLI with a browser workspace, or headless with `openscience run`, always against a project folder on your machine. The default Research agent plans the task (or agrees a plan first with `/plan`), searches web and scholarly sources and registered scientific databases such as UniProt, ChEMBL, PubMed, and arXiv, writes and executes code, and saves figures, reports, and the code that produced them. Every turn shows what it thought, searched, ran, and wrote. It can hand bounded work to workers: a read-only explore scout, ML, biology, physics, chemistry, and data specialists, and a general worker. Hundreds of bundled skills, most adapted from open collections such as K-Dense's Scientific Agent Skills and Orchestra Research's AI Research Skills, add domain procedures, and experimental adapters call ten NVIDIA BioNeMo NIM models with your own NVIDIA key. For heavier jobs it dispatches approved runs to Modal GPUs, SSH hosts, or Slurm and PBS clusters, and Autoresearch runs a metric-driven loop of experiments with a baseline, an idea queue, kill criteria, and a ledger of what was kept. The workbench itself is free. You pay for models through your own provider keys, a supported ChatGPT sign-in, a local model via Ollama or LM Studio, or Ace, Synthetic Sciences' pay-as-you-go managed access with no subscription. The tradeoffs: first-run setup needs a Synthetic Sciences account even when you use your own keys, session traces are shared by default while signed in, many skills need extra software or accounts before they work, and the project is young, with near-daily releases since its repository was created in July 2026.
Are you the founder? Claim this listing →