snowcite

MCP server for systematic literature review. Search arXiv / Semantic Scholar / OpenAlex / Crossref / PubMed, snowball through citations, review papers through chat, keep a notes graph and cross-paper synthesis. Writing the document is left to the model and its skills — snowcite builds no PDF.

No web UI, no browser, no second terminal. Claude reads abstracts in batches, pre-filters autonomously, surfaces borderline cases to you in chat.

How it works

flowchart TD
    Start([User: I want a systematic review]) --> Init

    subgraph Phase1["Phase 1 — Onboarding"]
        Init["init_project<br/>discipline, standard, backend, language"]
        Init --> GenFiles["Generate<br/>CLAUDE.md<br/>.claude/settings.json<br/>.claude/agents/*"]
        GenFiles --> Criteria["review_criteria<br/>include / exclude / categories"]
    end

    Criteria --> Search

    subgraph Phase2["Phase 2 — Search &amp; review"]
        Search["search_papers<br/>auto_save, no abstracts in context"]
        Search --> ReviewLoop

        subgraph ReviewLoop["Batch review loop"]
            direction TB
            GetBatch["papers status=unreviewed<br/>limit=20, no abstracts"]
            GetBatch --> AutoClassify{"Claude<br/>classifies"}
            AutoClassify -->|clear| BulkApprove["review_batch<br/>auto_high"]
            AutoClassify -->|borderline| AskUser["Show to user<br/>facts only, no recommendation"]
            AskUser --> UserDecide["User: i / e / m"]
            UserDecide --> SingleStatus["review_batch<br/>reviewed_by=user"]
            BulkApprove --> Summary["review_summary<br/>≤500 words, rolling"]
            SingleStatus --> Summary
            Summary --> GetBatch
        end

        ReviewLoop --> Snowball{"Deeper?"}
        Snowball -->|yes| Expand["snowball<br/>references + citations"]
        Expand --> Search
        Snowball -->|no| Outline
    end

    subgraph Phase3["Phase 3 — Structure"]
        Outline["bulk_create_sections<br/>title + scope per section"]
        Outline --> OutlineOk{"User<br/>approves?"}
        OutlineOk -->|edits| Outline
    end

    OutlineOk -->|yes| Phase4

    subgraph Phase4["Phase 4 — Write by section"]
        Draft["update_section<br/>draft"]
        Draft --> Critique["get_section_critique_inputs<br/>academic-reviewer subagent"]
        Critique --> Record["record_section_critique<br/>stop: no blockers or 2 iterations"]
        Record -->|blockers| Revise["revise_section"]
        Revise --> Critique
        Record -->|stop| NextSec{"More<br/>sections?"}
        NextSec -->|yes| Draft
    end

    NextSec -->|no| Done([Sections drafted — writing skills take over])

    classDef user fill:#fef3c7,stroke:#d97706,color:#000
    classDef auto fill:#dbeafe,stroke:#2563eb,color:#000
    classDef agent fill:#fce7f3,stroke:#db2777,color:#000
    classDef output fill:#dcfce7,stroke:#16a34a,color:#000

    class AskUser,UserDecide,OutlineOk,FixLang user
    class AutoClassify,BulkApprove,Search,Summary,Outline,Draft,Record,Compile auto
    class Critique,Humanizer agent
    class Done,GenFiles output

Legend — 🟨 user decisions, 🟦 automatic MCP calls, 🟪 subagents (fresh-context review), 🟩 outputs.

Install

claude plugin marketplace add notaskynet/snowcite
claude plugin install snowcite@snowcite-plugins           # MCP server + review / synthesis / section skills
claude plugin install academic-writing@snowcite-plugins   # optional: Russian style, critic, GOST 7.32

Server only, without skills:

claude mcp add snowcite -- uvx snowcite
# optional keys for higher rate limits:
#   -e SNOWCITE_SEMANTIC_SCHOLAR_API_KEY=xxx -e SNOWCITE_OPENALEX_EMAIL=you@example.com

First run

> I want a literature review on X for my bachelor thesis, in Russian, GOST 7.32

Claude calls next_step(), collects the project profile (language, discipline, standard, optional author / institution / targets) and runs init_project, which scaffolds .snowcite/ in the current directory. From there next_step() drives the phases: criteria, search, review, low-confidence confirmation, synthesis, then outline and drafting with critique. Nothing but .snowcite/ is written to your project.

Projects live in directories

Like .git, a .snowcite/ subdirectory marks a project. Work on multiple reviews in parallel by cd-ing between them — no global state, no switching commands. git clone moves a project to a new machine.

my-thesis/
├── .snowcite/
│   ├── papers.db         # project DB — metadata, papers, notes, sections, artifacts
│   └── cache/            # scratch space, always gitignored
└── ...                   # your own files; the document is built with your own toolchain

Skills

The server enforces the gates (summary before the next batch, notes for approved papers, confirmed low-confidence calls); the skills carry the procedure.

Docs

Full documentation: notaskynet.github.io/snowcite. Coverage:

• Installation on macOS / Linux / Windows
• Workflow in depth (review loop, snowball, synthesis, section critique)
• Troubleshooting — safety refusals, source rate limits
• Artifacts — interviews, code and notes as sources

Development

git clone https://github.com/notaskynet/snowcite.git
cd snowcite
uv sync --group dev
uv run pytest
uv run ruff check
uv run ruff format
claude plugin validate .   # marketplace + both plugins

CI runs ruff + pytest on every push and PR.

What this isn't

• Not a web UI. Review happens in chat. Architectural choice, see CLAUDE.md.
• Not a Zotero replacement. We import BibTeX / RIS but don't integrate with Zotero's API.
• Not a full-auto thesis generator. Systematic review requires human judgment at criteria-setting, borderline papers, outline structure. snowcite automates the mechanical parts and keeps you in the loop for the rest.

License

MIT — see LICENSE.