Search literature, build traceable evidence, analyze data, write drafts, verify citations — and keep a human in control of every major research step.
Quick start · See it work · Research packs · Architecture · Examples
Evidence, not memory — factual claims are mapped back to literature or project data; unsupported claims are surfaced explicitly. Verification built in — citation metadata, identifier conflicts, and unsupported claims are checked by machine, not vibes. Human-controlled workflow — artifacts are saved to disk before each major stage is approved, revised, or rejected.
A failed search is a source gap, not a scientific conclusion.
A general coding agent can already search, code, and write. What it cannot do out of the box:
| Problem | What openscience does |
|---|---|
| Research is not one chat — it's a stateful, multi-stage process | A lifecycle router with persisted stages (question → literature → hypothesis → experiment → analysis → writing) and resume-from-disk |
| Claims in drafts drift away from sources | An explicit evidence chain: retrieval → EvidenceItem → whitelisted synthesis → draft → claim support check |
| Agents run end-to-end without asking | Stage gates: each lifecycle stage writes its artifacts first, then pauses for approve / revise / reject |
| Chinese research workflows are second-class everywhere | CNKI bibliography import, Wanfang API, GB/T 7714 citations, and an epidemiology/public-health pack |
You:
"调研 2023–2026 年 LLM agent 做蛋白 binder design 的进展，
找出主要方法、失败模式和可验证的研究空白。"
openscience:
1. searches OpenAlex / Crossref / arXiv / Wanfang, dedupes into papers.json
2. reads full text where available — everything else is honestly marked abstract-only
3. maps claims ↔ evidence (EvidenceItem with verbatim quotes + page anchors)
4. synthesizes an 8-field literature survey
5. writes the draft (GB/T 7714 + references.bib)
6. verifies every BibTeX entry via bibverify MCP
7. flags unsupported claims and retrieves missing evidence
8. stops — you approve, revise, or reject
Everything lands on disk:
papers.json · evidence.json · survey.md · draft.md · references.bib
citation report · .openscience/provenance.jsonl
Requirements: Python 3.11+, uv, and Claude Code.
/plugin marketplace add Hylouis233/openscience
/plugin install science-core@openscience
/plugin install science-literature@openscience
/plugin install science-verify@openscience
Then just ask in natural language:
帮我调研过去三年蛋白质 binder design 中 AI agent 的进展，
重点比较自动化设计流程、实验验证和失败案例。
openscience routes the request through literature-search → paper-read → literature-survey → review-writing → citation-verify automatically. First run, it will ask you to complete a short cold-start-interview to build your research profile (field, data sources, compute, writing style).
Checked: whether references resolve (DOI/PMID/arXiv), whether returned metadata matches the entry, whether a factual claim has a supporting source, whether the environment and actions are logged.
Not claimed: scientific truth. no_match means not found in the queried sources — never "fabricated". Verification does not prove a claim; it proves the citation trail is real, consistent, and reviewable.
| Pack | Install if you need… |
|---|---|
| science-core | The workbench: research profile, lifecycle routing, stage gates, provenance, reviewer |
| science-literature | Search, read, survey, and write literature (incl. Chinese sources) |
| science-verify | Citation verification (bibverify MCP), claim checks, evidence loop |
| science-compute | Python/R analysis, SSH/HPC/Slurm, long-task runs |
| science-data | Domain database connectors with a fail-closed license gate |
| science-epi | Epidemiology & public health (outbreak curves, SEIR, spatial, writing) |
Recommended starting set: science-core + science-literature + science-verify.
question ─► [gate] ─► literature ─► [gate] ─► hypothesis ─► [gate]
─► experiment ─► [gate] ─► analysis ─► [gate] ─► writing ─► reviewer
When using the full lifecycle, each stage is persisted and paused for human approval before progression. revise "feedback" reruns the stage with your note injected — old artifacts are archived, never overwritten. Individual skills can also run standalone.
- CNKI: no scraping, no captcha bypass — you export a bibliography from cnki.net,
cn-literatureparses it into the shared paper schema - Wanfang: official open-platform API (
WANFANG_TOKEN; degrades cleanly when unset) - GB/T 7714-2015 reference formatting alongside BibTeX
- science-epi: outbreak investigation, SEIR modeling, spatial epi, and Chinese public-health writing templates
- Missing data stays missing. No full text →
abstract-only. No database access → source gap. No citation match →no_match, not "fabricated". - Never turn tool failure into evidence. Every external call degrades to a structured state, visible in the output.
- Humans decide at gates. The agent prepares; you approve.
- Append-only provenance.
record_run.pylogs actions and environment fingerprints to.openscience/provenance.jsonl.
examples/demo-workspace— standard research workspace layout with sample artifactsexamples/demo-epi— a fully worked outbreak analysis (synthetic data): linelist → epi curve → SEIR → report
Design contracts (review fence, stage gates, slug contract, evidence capsule levels) are documented in docs/architecture.md. Release and contribution workflow: docs/publishing.md.
openscience is an independent implementation informed by several open-source research-agent projects. See THIRD_PARTY_NOTICES.md for design references and license attribution — including Hylouis233/bibverify, the citation-verification MCP.
MIT — Copyright (c) 2026 Hylouis233 and contributors.