Bring papers, code and research evidence into one local AI workbench.
Quick start · Workflow · Architecture · Contribute · Releases
English · 简体中文
Research moves between papers, terminals, notebooks and chat windows. Superscience puts those tools in a shared project workspace and keeps artifacts connected to their sources. Your project files and session database stay on your machine; model calls and scientific searches use the providers you configure.
Start without an API key: choose Open local tools to explore files, scientific search and notebooks. AI chat stays disabled until a provider and model pass verification. The included research starter gives you a reproducible first project.
Research preview: the project is available as source. See the validation record for executed checks and the release process for distribution requirements. A supported connector or build target does not by itself establish live-service or signed-desktop acceptance.
| Bring together | What you can do |
|---|---|
| Research and sources | Search a catalog of 51 scientific connectors spanning literature, datasets, physics, astronomy, materials, robotics and life sciences; inspect returned metadata and source links. |
| Code and computation | Work with project files, a real PTY terminal, stateful Jupyter notebooks and language-server diagnostics. |
| Models and tools | Configure model providers, custom agents, tool permissions, Skills and MCP connections. |
| Results and provenance | Save artifacts, inspect provenance records and export or replay selected research bundles. |
| Browser and desktop | Use the React workbench through a local Go server or the Wails desktop shell. CLI and ACP entry points share the backend runtime. |
| Optional team infrastructure | Run the separate self-hosted Cloud service with OIDC, PostgreSQL and S3-compatible storage. |
The connector count describes registered implementations. External services have their own availability, credentials, rate limits and terms. Superscience returns source failures explicitly instead of filling in fabricated results.
Actual local-tools workflow: a real Python kernel analyzes ten explicitly synthetic measurements. The model provider was unavailable, so AI chat remains disabled. Run this example.
Live literature search and a local share replay
Crossref returned 20 records for “robot learning”; the selected record was saved with its source DOI.
The selected artifact was published to the local server and its replay verified. This is not a hosted Cloud deployment.
YouTube · Bilibili · X · Douyin · Xiaohongshu profile · Video source
Download the 1080p / vertical films, subtitles and covers — source and media prerelease; no signed desktop installers.
- Investigate. Choose Open local tools, open a project, then search a scientific source. Inspect the record and its source URL before saving it as an artifact. Verify a provider later when you want AI assistance.
- Execute. Use the notebook or terminal to analyze your material. Review permission requests before a network, write or execution action proceeds.
- Trace. Keep outputs with their provenance, select the artifacts to share, and export a bundle for inspection or replay.
Architecture illustration; see the component map and data boundaries.
Install Go 1.26.10 or newer, the minimum declared in go.mod, plus Node.js 24 and pnpm 11.7.0, then run:
git clone https://github.com/eust-w/superscience.git
cd superscience
pnpm install --frozen-lockfile
pnpm --dir frontend build
go run ./cmd/superscience serveOpen http://127.0.0.1:4096/ and choose Open local tools. Open examples/research-starter as a project to browse the CSV and run its Python or notebook example. Its 10 synthetic measurements produce a mean latency of 125.20 ms; these are invented demonstration values, not a performance benchmark.
For AI chat, choose a provider, enter an accessible model ID and API key, and complete verification. Provider usage may incur normal charges. The UI currently stores credentials through macOS Keychain; Linux and Windows can use the environment-backed CLI path for provider execution while native credential-store support remains a roadmap item. Local tools do not require that credential setup, but scientific search still requires network access.
For Python notebooks, install the worker and a kernel before starting the server:
python3 -m venv .venv
.venv/bin/python -m pip install -r workers/python/requirements.txt
.venv/bin/python -m ipykernel install --prefix "$PWD/.venv" --name python3
export SUPERSCIENCE_PYTHON="$PWD/.venv/bin/python"
export JUPYTER_PATH="$PWD/.venv/share/jupyter"
go run ./cmd/superscience serveSee Getting started for isolated data directories, frontend development, R/Julia kernels, desktop builds, CLI usage and troubleshooting.
Superscience is a workbench with permission controls. Approved notebooks, shell commands, MCP servers and language servers run with your OS user's privileges. Use a container or VM when you need process isolation; read SECURITY.md before handling untrusted material or deploying Cloud.
pnpm --dir frontend exec playwright install chromium
make verifyThe verification target covers Go race tests, vet and vulnerability scanning; frontend types, tests and build; browser acceptance tests; a production dependency audit; and the production-placeholder scan. Browser E2E uses a named localhost protocol fixture, which tests integration without claiming real-model quality. Current outcomes and remaining limits are recorded in validation notes.
Launch verification includes actual Jupyter execution and a live Crossref search. The attempted provider check returned HTTP 503, so no successful external-model response is claimed. These are separate acceptance results.
Useful next contributions include native credential storage on Linux/Windows, broader live connector acceptance, complete workspace localization, and signed desktop distribution. These are work areas, not claims of shipped functionality or promised dates.
Start with a reproducible bug, a documented connector, a tested permission boundary or a clearer example. Read the contribution guide, report a bug or propose a feature. If the workbench is useful to you, a star helps others discover it.
| Guide | Contents |
|---|---|
| Getting started | Setup, provider credentials, notebooks, CLI and troubleshooting |
| Architecture | Components, request flow, storage and extension points |
| Contributing | Development loop, focused verification and review expectations |
| Security | Permission boundaries, private reporting and deployment guidance |
| Releasing | Packaging, checksums, signing and publication |
| Validation | Recorded checks and explicit acceptance boundaries |
Original repository code and original demo assets use Apache-2.0. Dependencies, models, datasets and external resources retain their own terms; see THIRD_PARTY_NOTICES.md.
The product boundary is inspired by OpenScience. Superscience is implemented with Go, Wails and React. The ignored reference/ directory is a local analysis aid and is not part of the distributed source.