OpenScience: Open-Source AI Workbench for Scientific Research
Give it a goal — it reads the literature, writes and runs code, runs experiments, and writes up the findings.
Evidence shows a permission system that prompts users, but explicitly states it is not a sandbox, and no detailed least-privilege mechanism is provided. User confirmation exists (permission prompts) but implementation details are not shown. Data flow transparency: README mentions sessions, skills, artifacts, and provenance stored on disk, but does not detail data flows. Sensitive data handling: mentions credentials are filtered and redacted, but no implementation details. Dependency security: has dependency overrides and patches, but no vulnerability scanning evidence. External effects: explicitly states agent can run shell commands and access web, but no restrictions. Rollback: not mentioned. Source attribution: has NOTICE file, but not detailed.
Self-consistency: README, SECURITY.md, and CI config are consistent on security model, but some descriptions (e.g., sandbox) may conflict with code. Dependency availability: dependencies published via npm and GitHub, but no availability guarantees. Failure messages: tests have error handling, but no user-visible failure message examples.
Audience and scenarios: clearly targets researchers, provides multiple scenarios. Capability boundaries: no explicit limits on agent capabilities. Trigger precision: no details on command or skill triggers. Environment fit: CI for Linux, macOS, Windows, but no detailed system requirements.
Information architecture: README provides clear directory structure. Install notes: provides npm install and npx commands. Naming stability: no mention of API stability. Examples and FAQ: has quickstart, but no FAQ. Known limitations: SECURITY.md clearly states no sandbox. License: Apache-2.0, complete. Versioning changelog: has CHANGELOG.md, but content not provided. Maintenance responsibility: has SECURITY.md and contributing guide, but maintainers not explicit.
Output usability: provides browser workspace, interactive output. Marginal value: provides 290+ skills and database integrations, clear value. Cost benefit: free with own keys, but no performance or cost data.
Claim traceability: README claims lack specific evidence. Cross-source corroboration: has CI and tests, but no independent verification. Fact-inference separation: not explicitly distinguished.
- Agent is not sandboxed; permission system is not an isolation boundary; run in container or VM.
- Dependency security lacks vulnerability scanning evidence; review dependencies yourself.
- Rollback mechanism not mentioned; operations may be irreversible.
What does this agent do, and when should you use it?
OpenScience is an open-source AI workbench for scientific research, providing a complete research environment within a browser. It integrates literature retrieval, code execution, experiment running, and paper writing. It is model-agnostic, supporting Anthropic, OpenAI, Google, and many other providers via your own API keys, with no account required. It includes specialized agents for research, biology, physics, and ML, plus 290+ skills covering training, evaluation, datasets, molecular biology, cheminformatics, and cloud compute. Optional Atlas managed platform adds curated models, a research graph, and cloud compute.
Users set a research goal in the browser workspace; the agent autonomously works through the research loop: reading relevant literature (via UniProt, PDB, Ensembl, ChEMBL, PubChem, arXiv, OpenAlex, etc.), forming hypotheses, writing and running code (using shell, editor, LSP, MCP servers, and other tools), running experiments on real compute, and writing up the result. It is extensible with custom agents, commands, tools, plugins, and a TypeScript SDK.
- ML researchers quickly exploring new ideas, letting the agent handle literature review and initial experiments.
- Biology researchers querying UniProt, PDB, and other databases and running molecular analyses.
- Physics or chemistry researchers using the agent for data analysis and simulation.
- Research teams comparing results across multiple providers (Anthropic, OpenAI, Google).
- Developers integrating the agent into existing workflows via the TypeScript SDK and MCP support.
What are this agent's strengths and limitations?
- Model-agnostic, supporting multiple providers with your own keys, no account needed.
- Built-in scientific database connectors (UniProt, PDB, ChEMBL, etc.) for direct querying.
- Full browser workspace with file tree, editor, terminal, and inline rendering for molecules and structures.
- Highly extensible with LSP, MCP, custom plugins, and a TypeScript SDK.
- Agent is not sandboxed; security risk, recommend running in a container or VM.
- Requires configuring API keys for each provider, potentially multiple accounts and costs.
- Linux kernel 5.1+ required; CentOS 7's 3.10 kernel unsupported, limiting deployment environments.
- Optional Atlas platform requires a prepaid wallet for its managed models, adding cost.
How do you install or deploy this agent?
- Ensure Node.js and Bun 1.3+ are installed. 2. Install globally via npm:
npm install -g @synsci/openscience. 3. Runopenscienceto open the workspace in your browser.
How do you use this agent?
- Set an API key from a provider, e.g.,
export ANTHROPIC_API_KEY=sk-ant-.... 2. Runopenscienceto start. 3. Select a model from the model selector, and optionally specify a project directory:openscience ~/code/my-project. 4. Set a research goal; the agent will execute the research process.