Eigma

app.eigma由 eigma.app 发布官方发布者v1.0.0更新于 Oct 24, 2026

Files what you learn into a personal wiki and quizzes you before you forget it.

已验证Streamable HTTP可网页运行效率与工作流知识与记忆

概览

AI 生成的概览

Eigma 把你在 AI 对话中学到的内容归档为个人 wiki，并在你遗忘前出题复习，通过托管 MCP 连接器使用。

可以做什么
  • 搜索和浏览 wiki，支持跨文字系统与页面别名的多语言匹配
  • 读取页面并据此作答，处理草稿、读取记忆状态、翻译页面
  • 对照真实来源核查页面，提出带脚注的改写建议
  • 把对话内容保存为页面，归入正确章节并附上测验题
  • 合并重复页面、重整章节结构，新增、替换或删除测验题
  • 列出你最快会遗忘的页面供复习
适用场景
适合希望把助手对话中的知识长期保存为可搜索的个人 wiki 并进行间隔复习的人，也适合让助手在后续对话中把这些笔记当作上下文。适用于已在 ChatGPT、Claude、Claude Code 或 Codex 中工作并想要记忆层的用户。
运行要求
托管在 my.eigma.app 的远程 MCP 服务器，通过 streamable HTTP 访问，ChatGPT 与 Claude/Claude Code/Codex 使用不同地址。需要先在 eigma.app 注册 Eigma 账号并通过 OAuth 登录授权；未声明 API 密钥或环境变量。Claude 需在每个对话中手动开启该连接器。
安装前请注意
登录后页面会写入你所授权的 Eigma 账号，请确认账号无误。写入工具会改动 wiki，但所有变更以可审阅的变更集形式提交，需逐条批准或丢弃。连接失败可能悄无声息，安装后应确认连接状态。公开仓库只包含工具描述和示例服务器，不含托管服务本身。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Eigma，将其添加到工作区。
  2. 连接 所需的 凭据。它们会被私密存储，绝不会在公开页面上展示。
  3. 为需要使用其工具的对话启用该服务。

Web executable，通过 Streamable HTTP。 远程服务在工作区中配置后即可从网页运行时运行。

交给智能体安装

把这段提示词粘贴到 Claude Code、Codex、Cursor 等智能体，或 SourceWeft 对话中。智能体会读取该服务器的安装指南，先向你展示发布者、端点和登录方式，经你同意后再把它添加到你的客户端。密钥不会写进任何命令。

阅读 https://sourceweft.com/mcp/app.eigma%2Feigma/install.md，并按其中说明把这个 MCP 服务器接入我的客户端。修改任何配置之前，先告诉我它的发布者、端点和登录方式，等我确认。不要把 API 密钥写进命令或我会提交的文件；问我密钥应从哪里获取。

查看智能体所依据的安装指南

其他 MCP 客户端

把它添加到你客户端的 mcpServers 配置中。 客户端会在首次连接时完成 OAuth 登录。

{
  "mcpServers": {
    "eigma": {
      "type": "http",
      "url": "https://my.eigma.app/api/mcp"
    }
  }
}

README

[Eigma]

eigma-mcp

Every tool description, input schema and shared extraction rule from the Eigma MCP connector, plus a runnable example server. The text a model reads, not the service behind it.

[License: MIT] [Node]

eigma.app · Docs

What Eigma is

Eigma turns your conversations with an AI assistant into a personal wiki you actually remember.

An assistant can explain something well, and almost none of it survives you closing the tab. A week later you remember that you understood the thing, but not the thing itself, and what you do recall is often subtly wrong. Eigma saves the part worth keeping, files it under the right chapter, and asks you about it again before you forget.

[The Memory State view: a rotating map of nodes, one per page, each glowing according to how well it is remembered]

Your knowledge as a map. Each page is a point of light, and its brightness shows how well you still remember it rather than how recently you opened it. Pages you have internalized stay bright; pages you are forgetting go dim.

Eigma connects over MCP, so it works inside the assistant you already use — ChatGPT, Claude, and coding agents like Claude Code and Codex. Your assistant does the thinking; Eigma keeps the result. The app is available in English, 中文 and Deutsch. → eigma.app

Connect it to your assistant

Setup takes about a minute, and you do it once per assistant.

Create your Eigma account first, at eigma.app. Partway through, you will be asked to sign in, and your pages will go to whichever account you sign in with.

ChatGPT — its own address 
https://my.eigma.app/api/mcp-chatgpt
  1. Settings → Plugins
  2. Turn on Developer mode. The option to add your own plugin does not appear until you do.
  3. Browse plugins → + to add one
  4. Name it Eigma and paste the address into the server URL field
  5. Sign in with your Eigma account and approve the connection
Claude, Claude Code and Codex — one shared address 
https://my.eigma.app/api/mcp

Claude: Customize → Connectors → + → Add custom connector, paste the address, Add, then sign in.

⚠️ Then open a chat, click the + at the bottom left, choose Connectors, and switch Eigma on. Claude uses only the connectors that are switched on in that particular conversation. If Eigma looks connected but nothing happens, this is almost always why.

Claude Code:

bash
claude mcp add --transport http --scope user eigma https://my.eigma.app/api/mcpclaude mcp login eigma

Codex:

bash
codex mcp add eigma --url https://my.eigma.app/api/mcpcodex mcp login eigma

The login command opens a browser window so you can sign in and approve. Afterwards, run claude mcp list (or /mcp in a chat) to confirm the connection: worth doing, because a failed connection fails silently.

Then, in any chat: "organize this into Eigma". Your assistant turns the conversation into pages, files each page under the right chapter, attaches a few questions to it, and sends the result to your review inbox.

Screenshots and troubleshooting: eigma.app/docs.

Menu paths verified against the ChatGPT and Claude apps in September 2026.

What you can ask it to do

Eight tools, five of them read-only. The read-only tools are what make your wiki usable as context inside the chat you are already in, rather than as something you have to open separately.

Say something likeWhat happens
"What do I have on transformers?"Matches across scripts, using CJK bigrams, German compound splitting and each page's aliases, so a page filed in English stays findable in the language you ask in
"Use my notes on Kubernetes to answer this"Reads the page and answers from it
"Check this page against real sources"Re-reads it, tests the claims, proposes a rewrite with inline footnotes and a References section
"These two pages are about the same thing, merge them"Merges the two: the absorbed page's body is appended to the page you keep. Useful against the near-duplicates a growing wiki accumulates
"Go through my Cloud chapter and tell me what is badly organised"Reads the chapter, then proposes all the moves and renames as a single changeset you can review in one pass
"These questions are weak, write better ones"Adds, replaces or removes specific quiz questions
"What should I review today?"Lists the pages you are closest to forgetting, rather than the ones you opened most recently
"Translate my page on embeddings into German"Sends a translated copy to your inbox; the original page is untouched

Everything that changes the wiki arrives as a reviewable changeset: a list of operations you approve or drop one by one. An assistant deciding that a change is a good idea is never enough to apply it.

What is in this repository

This repository holds everything a model reads when it talks to Eigma: the tool descriptions and the input schemas. The connector itself is a hosted MCP server with OAuth, and that part is not here.

Take what is useful: vendor src/rules.ts and compose your own descriptions from it, copy the ingestSchema.ts pattern, or fork example/server.ts as a starting point.

bash
npm installnpm testnpm run example    # boots the example server and prints its address
src/toolSpec.tsDescriptions for the six read tools (browse, search, read a page, resolve drafts, read memory state, translate), plus the outline and search helpers behind them.
src/writeSpec.tsDescriptions for the two write tools, each in a full and a slim variant. These are the tools that change your wiki, so their wording matters most.
src/rules.tsThe 20 shared extraction rules that every tool description and the server instructions are composed from. Change a rule here and the change takes effect in every channel at once.
src/ingestSchema.tsThe save tool's input schema.
src/changeOpsSchema.tsThe edit tool's operation vocabulary, typed.
src/skill.tsThe connector's skill and its server instructions: when to act, what order to call tools in, how not to loop.
src/renderSurface.tsA way to drive your own handler and get back the exact bytes tools/list puts on the wire.
src/tools.tsThe ordered tool registry: reads before writes, plus the per-host type that decides which variant of each description is used.
src/textNormalize.tsThe single text normalizer that every cross-language comparison in this repository uses.
skill/eigma-filing/The generated Skill package, for hosts that load skills. Rebuild with npm run build:skill.
example/server.tsA working MCP server over a Map. No database, no auth, ~330 lines; six of the eight tools wired end to end.

License

MIT © Eigma. The Eigma name, logo and screenshots are not covered by the MIT grant.

来源：README.md，提交 c6ed777

工具

0
该服务器需要登录，因此 SourceWeft 无法在没有账号的情况下读取其工具，也未收录。已连接账号所能看到的内容可能因账号及其权限而异。

版本历史

1
  1. v1.0.0最新Sep 16, 2026

评分

已满足适用检查项的 85%

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 天内可连通必需25 分
  • 已声明认证方式必需10 分
  • 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