OpenScience¶
external · status: active · focus: end-to-end · discipline: general · started: 2026
Project page: https://github.com/synthetic-sciences/openscience
Source: projects/landscape/openscience.yml
Positioning¶
An open-source (Apache-2.0), model-agnostic AI workbench for scientific research, launched 2026-07-03 as an explicit open alternative to Claude Science: a Bun/TypeScript monorepo whose CLI starts a local server hosting a browser workspace, an agent runtime (default research agent plus biology/physics/ml specialists with critique and literature-review sub-agents), 292 bundled markdown skills, and ~30 scientific-database connectors (UniProt, PDB, ChEMBL, PubChem, arXiv, OpenAlex, Semantic Scholar). Sits in the big-lab workbench layer alongside Claude Science, Prism, and Google Co-Scientist, but self-hosted and bring-your-own-key.
Distinctive contribution¶
The first credible open-source counterweight to the closed research workbenches: per-request model routing across 75+ providers, local-first operation with no account required, and the entire stack — prompts, agents, skills, connectors, SDK, plugin runtime — auditable and editable in one repo, with an optional managed platform (Atlas) kept strictly non-required. Ships an RSI trajectory critic (correctness / efficiency / coverage / reproducibility, 0-100) and an optional blind-reviewer gate at session finalize.
Evaluation scores¶
| Dimension | Score (0–3) | Note |
|---|---|---|
| Lifecycle coverage | 2 | Eight stages from hypothesis through write-up; the peer-review skill and blind-reviewer gate are optional runtime checks, not a declared referee-simulation or dissemination stage. |
| Autonomy level | 2 | Supervised agent: user sets a goal, the agent runs the research loop with tool permissions and a read-only plan mode; user reviews the resulting artifacts. |
| Architectural transparency | 3 | Full code, agent prompts (per-provider prompt files), skills, connectors, config schema, ARCHITECTURE.md and AGENTS.md all public; released builds fetch the skill catalog from the Atlas index, source builds bundle it. |
| Inputs supported | 3 | Goal prompts, existing project directories, and local datasets; literature APIs (arXiv, OpenAlex, Semantic Scholar, Europe PMC) plus ~30 scientific databases, with private local data by construction. |
| Outputs / reproducibility | 2 | Sessions, artifacts, and provenance persist on disk and are shareable as links, and the critic scores reproducibility — but end-to-end re-runnability from declared inputs is not demonstrated. |
| Internal evaluation | 1 | Demo sessions and launch write-ups only; no benchmark results or third-party evaluation of research-output quality published. |
| Openness | 2 | Apache-2.0, npm-installable, BYOK free and non-gated; demo examples need paid provider keys and heavier experiments need cloud compute, so reproducibility on commodity hardware is partial. |
| Maturity / traction | 1 | 2,696 stars / 374 forks in under three weeks amid a coordinated 'open Claude Science alternative' press cycle; velocity is plausible for the niche and no direct evidence of star manipulation was found (star-timestamp audit unavailable), but the core team is 2-3 people and v1.3.4 landed eight days after repo creation — too young to call durable adoption. |
| Cross-family policy | 1 | Per-request routing across 75+ providers makes cross-family executor/reviewer setups configurable (e.g., critique sub-agent on another family), but cross-family review is neither default nor required. |
| Runtime assurance | 1 | Optional blind-reviewer gate at finalize plus RSI trajectory critic and tool-permission prompts; all optional, with no mandatory claim-audit or citation-verification gating. |
| Cross-platform portability | 2 | 75+ model providers and macOS/Windows/Linux via npm or platform binaries, but a single self-contained runtime — not deployable across other agent frameworks. |
Scored on 2026-07-23. See the evaluation rubric.
Tags¶
Pipeline stages: hypothesis-generation literature-discovery literature-synthesis data-acquisition data-analysis code-generation paper-drafting revision-editing
Architectural features: multi-agent tool-use iterative-loop persistent-memory artifact-versioning
Inputs: research-goal-prompt project-directory user-datasets
Outputs: analysis-artifacts experiment-code figures manuscript-drafts
Data sources: scientific-databases user-provided
Knowledge sources: arxiv openalex semantic-scholar curated-skill-library
Limitations¶
- Near-twin disambiguation: ai4s-research/open-science (~900 stars, created the same day, since rebranded 'Open Science Desktop') is not a fork or mirror but an independently built Tauri desktop app on an attributed OpenCode sidecar, from an anonymous org registered 2026-06-27 with bot-named commits and the name-adjacent domain openedscience.com; org identity (Synthetic Sciences, est. 2025, named maintainers), npm distribution (@synsci/openscience), the openscience.sh docs domain, and launch press mark synthetic-sciences/openscience as the canonical OpenScience — but the namespace confusion persists.
- Runtime provenance is under-disclosed: the agent core closely mirrors the MIT-licensed OpenCode CLI (identical provider-prompt file set — anthropic.txt, beast.txt, gemini.txt, qwen.txt — and session-module layout), yet the NOTICE file credits no OpenCode lineage.
- Three weeks old with a 2-3 person team and no published evaluation of output quality; by the project's own security note the agent is not sandboxed — the permission system is not an isolation boundary.