CiteVahti

io.github.heidihelena由 GitHub 账号 heidihelena 发布v0.46.0更新于 Oct 24, 2026

Checks whether each cited source supports the claim it backs. Blinded human-first, local-first.

已验证STDIO仅桌面效率与工作流知识与记忆

概览

AI 生成的概览

本地优先的引文支撑审查工具,让助手协助预筛稿件中的论断,而由用户本人逐条评分并做最终决定。

可以做什么
逐条梳理稿件中的论断,助手可建议并暂存“所引文献是否支撑该句”的检查,但不能代替用户评分或决定。提供盲评的 AI 第二意见,在用户完成评分前始终隐藏。把每项决定保存为用户电脑上带时间戳、防篡改的审查记录。建立可复用的私人证据地图,记录以往的论断判断,并将审查记录导出为 Word 文档、可打印报告或审查资料包。可选地把结果写回稿件或 Zotero,且总是先预览并确认。
适用场景
适合研究人员、临床医生和学生,在投稿或送审前确认论文中每条引文确实支撑其对应句子,并希望留下可供导师、合著者或期刊核查的审查记录。
运行要求
作为本地进程运行在用户机器上(PyPI 包 citevahti,或 Claude Desktop 扩展包)。代理服务器默认关闭,需从应用或菜单栏启动。无需账号或 API 密钥;可选的 AI 第二意见可使用本地 Ollama 模型或用户自己的 OpenAI/Anthropic 兼容 API 密钥。桌面应用仅支持 macOS,Windows 和 Linux 使用 Claude Desktop 扩展。
安装前请注意
可选的 API 密钥 AI 模式会把论断及论文标题和摘要发送给用户配置的服务商。对外请求还可能访问文献服务(PubMed、OpenAlex、Semantic Scholar、Crossref)、Zotero、PyPI 版本检查、Pandoc 下载以及可选的 RFC 3161 时间戳。对稿件或 Zotero 的写入需预览和确认,且可撤销。它只检查引文支撑,不验证论断真伪,也不认证稿件质量。

安装

在 SourceWeft 中

  1. 打开 控制台中的 CiteVahti,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

CiteVahti

Does the cited paper actually support the claim? Check every citation in your manuscript — before a reviewer does.

[CiteVahti workflow: claim, source evidence, human rating first, blinded AI second opinion, agreement or adjudication, and an audited decision — with local manuscript, previewed writes, Zotero, undo and a hash-chained record as the supporting rail]

CiteVahti walks you through your manuscript one claim at a time. You decide whether each cited source really backs the sentence; an optional AI gives a blinded second opinion (you rate first, it can never overrule you); and every decision is saved as a timestamped, tamper-evident record on your own computer.

[The CiteVahti review panel — your manuscript on the left, the evidence-and-decision card on the right.]

Free while in beta. CiteVahti is local-first: your manuscript stays on your device unless you choose to turn on an external AI model. No account. Nothing uploaded.

What beta means here: the review workflow and your data format are stable, but the tool is young and your feedback still shapes it. The app does not update itself yet — when a new version ships you replace it yourself (steps under Updating, below; the panel can check for a newer version, but only if you turn that on). If something is confusing or judged wrong, say so — beta reports are how it improves.

Who it's for

Researchers, clinicians, and students who write papers and want to be sure every citation holds up — without learning new software or touching a command line. If you can open a file on your computer, you can use CiteVahti.

What it does for you

  • Catches citations that don't support the sentence they're attached to — the kind a reviewer or editor will flag.
  • Keeps you in charge. You make every call; the AI is only ever a second opinion, shown after you've decided.
  • Leaves a clean paper trail — a timestamped record of what you checked and concluded, ready for a supervisor, co-author, journal, or registry.
  • Builds a reusable evidence map so you don't re-do the same checks next project.
  • Stays private by default — everything lives in a folder on your Mac or PC.

Get started in 2 minutes — no terminal, ever

On a Mac, use the desktop app. It's a normal app: download it, open it, review.

  1. Download citevahti-<version>-macos-arm64.app.zip from the latest release.
  2. Unzip it and drag CiteVahti.app into your Applications folder, then open it like any other app. (It's signed and notarized — it opens with no security warnings.)
  3. Choose a folder to keep your reviews in when it asks.
  4. Drag your manuscript (.docx or .md) onto the window. The panel walks you through the rest, one step at a time — there's nothing to memorise.

Want an AI assistant to help pre-screen citations? Whenever you're ready, click the CiteVahti menu-bar icon → Start Agent Server. CiteVahti first explains exactly what an assistant can and cannot do — and either way, you rate every claim yourself and make every final call.

Or: use it inside Claude Desktop

If you already work in Claude, install CiteVahti as a one-click extension and drive the review by chatting. (This is also the path on Windows and Linux, where the desktop app isn't available yet.)

  1. Download the extension for your computer from the latest release:
    • macOS (Apple Silicon) — citevahti-<version>-macos-arm64.mcpb (signed + notarized)
    • Windows — citevahti-<version>-windows-x64.mcpb
    • Linux — citevahti-<version>-linux-x64.mcpb
  2. Double-click the downloaded file to add it to Claude Desktop.
  3. Choose a folder for your reviews when asked.
  4. Tell your assistant: "Check the claims in my manuscript using CiteVahti."

First time on macOS you may see a one-time security prompt — the extension is a zip, which macOS can't pre-stamp even though the app inside is signed and notarized. If it appears: right-click the file → Open, or go to System Settings → Privacy & Security → "Open Anyway". After that it installs and runs normally.

First time on Windows you may see a SmartScreen warning ("Windows protected your PC") — the Windows build is unsigned, so Windows can't check who published it. If it appears: click More info, then Run anyway. After that it installs and runs normally.

Updating — your review data is never touched; only the program is replaced
  • Desktop app: quit CiteVahti.app, download the newest citevahti-<version>-macos-arm64.app.zip from the latest release, and replace the old CiteVahti.app in Applications with it. Your reviews live in the project folder you chose, not inside the app, so nothing is lost.
  • Claude Desktop extension: remove the existing CiteVahti extension, fully quit and reopen Claude Desktop, then add the newest .mcpb from the latest release. Ask your assistant to run the status tool to confirm the running version. (macOS Intel: install via pip install "citevahti[mcp]", or run the Apple-Silicon .mcpb under Rosetta.)
  • pip: pip install --no-cache-dir --upgrade "citevahti[mcp]", then pip show citevahti to confirm the version.
For developers & technical readers — pip, CLI, architecture, security

Everything below assumes a terminal; none of it is needed to use the app.

Try it on synthetic data (nothing of yours touched — builds an invented ledger and opens the panel showing every claim state):

bash
pip install "citevahti[mcp]"citevahti demo

Run the review panel from the terminal:

bash
pip install "citevahti[mcp]"citevahti run

CiteVahti opens the review panel on your machine at 127.0.0.1 and tells you the one next thing to do.

Run the desktop app from source (the same supervised shell as the packaged .app):

bash
pip install "citevahti[app]"citevahti-app

What the app actually runs: CiteVahti.app is a thin shell supervising two local processes — the review panel engine (for you) and an optional MCP agent server (for a chat client; it can suggest and stage, but it can never rate for you, see your rating first, or make the final call). Both bind only to 127.0.0.1, write rotating logs (menu-bar icon → Open Logs Folder), exit on their own if the app dies, and shut down cleanly when you quit. The agent server is off until you enable it and can be stopped anytime from the menu bar.

Deeper reading:

How it works

CiteVahti always tells you the one next thing to do — there's no command to remember and nothing to set up.

[A "what's next" banner above the manuscript names the next action and takes you to it.]

You start by adding your manuscript — choose a Word or Markdown file, drag it onto the window, or paste the text. Nothing is uploaded.

[Start your review: choose a file, or paste the text instead.]

Then you work through each claim on a simple card that walks a visible Rate → Reveal → Decide → Write path, so you always know where you are:

  1. Rate — you judge whether the cited paper supports the claim.
  2. Reveal — now the optional AI second opinion appears (never before).
  3. Decide — accept, caution, review, or reject. The decision is yours.
  4. Write — optionally update your manuscript or Zotero, always after a preview.

[The review card showing the Rate → Reveal → Decide → Write steps.]

A record you can hand to anyone

Every decision is saved to a timestamped, tamper-evident review record — a plain-language log of what you reviewed and in what order. Export it as a Word document, a print-ready report, or a review-packet .zip (which even includes an auto-filled methods paragraph with your review's real numbers) for a methods section, a co-author, a supervisor, or a journal.

[The Output section: export your review record as a .zip, Word, PDF, or Markdown.]

Build an evidence map you can reuse

Stop losing evidence work after each project. Most citation checks vanish after one manuscript or one deadline. Every claim you assess in CiteVahti is saved to your own private evidence map, so next time you work on the same topic you don't start from zero — pull back earlier judgments, see what you already checked, and spot where the literature is thin, contradictory, or overclaimed.

It stays on your machine. Sharing into the wider, de-identified community evidence map is a separate, opt-in step, and you preview the exact payload before anything is sent — never your manuscript, search history, patient data, registry data, or private project identifiers.

How your data is handled →

The AI is a second opinion — never the judge

You always rate first. CiteVahti can then show a second rating from an AI, and compare the two — but the AI value stays hidden until your rating is in, and the final decision is always yours. There's no hidden AI subscription: pick the mode you want, or leave AI off entirely.

[AI second opinion settings: Off, local AI, or your own API key.]

You can get the second opinion from:

  • your MCP assistant (e.g. Claude Desktop),
  • a local model on your own computer (free, fully private — see below), or
  • your own OpenAI/Anthropic-compatible API key.
Run the AI on your own machine (free, nothing leaves your device)
  1. Install Ollama (macOS: brew install ollama).
  2. Pull a model — qwen2.5 is recommended for claim checking:
    bash
    ollama pull qwen2.5
    (Pick a smaller tag if your machine is tight on memory, e.g. ollama pull qwen2.5:3b.)
  3. In the panel, open Settings → AI second opinion → Local AI. CiteVahti detects the installed model and pins its exact version so the rating stays auditable.

To update later, re-pull the same name (ollama pull qwen2.5).

Privacy Policy

CiteVahti is local-first. Full policy: https://vahtian.com/citevahti/privacy (source: docs/PRIVACY.md). In short:

  • No telemetry, no analytics, no account — CiteVahti collects nothing about you and nothing for us, and never phones home.
  • Your manuscript, ratings, and review record stay in a folder on your computer.
  • The panel runs only on your own machine (loopback, 127.0.0.1).
  • AI is optional and blinded; local and bring-your-own-key modes are supported.
  • Changes to Zotero or your manuscript always require a preview and your confirmation, and can be undone.
  • Outbound calls are enumerated, not hidden. The main ones: the literature services you choose to search (PubMed, OpenAlex, Semantic Scholar, Crossref), your Zotero if you connect it, the optional AI second opinion (off by default; API mode sends the claim plus the paper's title and abstract to the provider you configure), a user-initiated check to PyPI for a newer version, a one-time Pandoc download if you ask for a Word export without Pandoc installed, and opt-in RFC 3161 timestamping (audit-head hash only). The complete table — destination, payload, trigger, default — is in docs/PRIVACY.md.

What CiteVahti writes

CiteVahti never silently changes Zotero or your manuscript. Every write follows the same gate:

Preview → Confirm → Audit → Undo available

What it does and does not check

CiteVahti checks citation support, not the truth of the underlying claims. It tests whether the cited paper backs the sentence — it does not certify that every claim in the manuscript was entered, and it is not a clinical or scientific oracle. Blinding is enforced by the workflow (the AI value is withheld until you rate) and recorded in the review record with timestamps, so the order is auditable, not assumed.

Responsibility & disclosure

CiteVahti is a local-first citation-support audit tool for manuscript claims — it helps authors test whether the cited evidence supports each claim before submission.

Using CiteVahti does not certify scientific truth, manuscript quality, publication suitability, or the absence of citation problems. It provides structured decision support and an audit trail; final responsibility remains with the human author, reviewer, editor, or institution. CiteVahti is developed by the author — any publication, review, teaching, or evaluation that uses it should disclose the tool use and the developer relationship where relevant. Ready-to-adapt disclosure text and the full statement are in docs/DISCLOSURE.md.

Learn more

Technical documentation (architecture, CLI, safety invariants, SBOM) lives under the "For developers" toggle above and in docs/.

Feedback & support

CiteVahti is a free beta and your feedback shapes it.

New here? Open the app and drag a manuscript in — or follow the Quickstart step by step.

Companion: FullVahti

FullVahti is a sibling Zotero plugin that finds free, legal open-access PDFs for your references and writes CiteVahti's check results back as tags — so the citation check and the full text live in one place. See the FullVahti README to install (a two-click Zotero plugin, no terminal).

License

Apache License 2.0 — see LICENSE and NOTICE.

来源:README.md,提交 a2655ae

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

1
  1. v0.46.0最新Sep 16, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分