ContextOverflow

org.contextoverflow由 contextoverflow.org 發布官方發布者v1.0.0更新於 Oct 24, 2026

Describe what's going wrong with your AI; get the named technique both of you can read. No auth.

已驗證Streamable HTTP可網頁執行知識與記憶AI 與機器學習

概覽

AI 產生的概覽

一個免費、免金鑰的遠端 MCP 伺服器，把描述的 AI 思考問題對應到具名脈絡工程技巧，並回傳可執行的鷹架。

可以做什麼
列出語料庫涵蓋的問題類別，以人類對 AI 行為的抱怨方式表述，並把症狀描述分類到相符的技巧，含糊時只提出一個釐清問題。依編號或名稱直接查找技巧。讀取技巧的機制、代理指令、敘述範本、驗證檢查與失敗模式。取得最小可執行鷹架，包含步驟、對使用者說的敘述句，以及證明其有效的檢查。
適用情境
當助手或其使用者想命名並修正反覆出現的工作問題（例如遺失脈絡、一味附和、遲遲不行動或做得比要求多），並希望取得附驗證步驟的結構化技巧而非泛泛的提示詞建議時適用。
執行需求
位於 contextoverflow.org 的遠端 Streamable HTTP 端點；未宣告驗證、API 金鑰或環境變數。需要網路存取。
安裝前請注意
伺服器無狀態且免金鑰，不會暴露憑證。回應包含供對使用者說出的敘述句與指令，apply_technique 回傳的鷹架可能被助手執行，執行前應先審閱。問題描述會傳送給第三方服務。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 ContextOverflow，將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Web executable，透過 Streamable HTTP。 遠端服務在工作區中設定後即可從網頁執行環境執行。

交給代理程式安裝

把這段提示詞貼上到 Claude Code、Codex、Cursor 等代理程式，或 SourceWeft 對話中。代理程式會讀取該伺服器的安裝指南，先向你展示發布者、端點和登入方式，經你同意後再把它新增到你的客戶端。金鑰不會寫進任何命令。

閱讀 https://sourceweft.com/mcp/org.contextoverflow%2Flibrary/install.md，並按其中說明把這個 MCP 伺服器接入我的客戶端。修改任何設定之前，先告訴我它的發布者、端點和登入方式，等我確認。不要把 API 金鑰寫進命令或我會提交的檔案；問我金鑰應從哪裡取得。

檢視代理程式所依據的安裝指南

其他 MCP 客戶端

把它新增到你客戶端的 mcpServers 設定中。

{
  "mcpServers": {
    "contextoverflow": {
      "type": "http",
      "url": "https://contextoverflow.org/mcp"
    }
  }
}

README

Context Overflow

Like Stack Overflow, but for context engineering — thinking techniques for humans and AI agents, from one corpus, through two doors.

Most guides teach you to write better prompts. Context Overflow teaches something different: how to work with AI so that it extends your thinking instead of replacing it — and gives your agent the same techniques through a protocol it can actually run.

The two doors

For humans — contextoverflow.org. A curated, categorized corpus of thinking techniques. No login, no feed. You arrive with a problem — the site opens with "What's going wrong?" — find the technique that fits, and learn to build it into a prompt yourself: worked, annotated examples, never copy-paste snippets. The research reference for what this counters: contextoverflow.org/cognitive-debt.

For agents — MCP at contextoverflow.org/mcp. The same corpus served over Streamable HTTP; free, keyless, stateless. Listed in the official MCP Registry as org.contextoverflow/library. Five tools:

ToolWhat it does
list_categoriesThe major problem categories, as a human would say them
classify_intentSymptom description → matching techniques; genuinely ambiguous → one clarifying question, never a guess
find_techniqueDirect lookup by name
get_techniqueMechanism, agent instructions, verification, failure modes
apply_techniqueThe runnable scaffold + narration line + the check that proves it worked

Every response carries a narration line, so the agent's technique use surfaces in-conversation in the same vocabulary its human learned on the site. When your agent says "one real unknown before I act," you know exactly which technique is running. Both sides of the pair get smarter; neither goes opaque. Setup for every client: contextoverflow.org/connect.

The major problem categories we've found so far

lost-the-thread · doing-my-thinking · confidently-wrong · agrees-with-everything · stalls-instead-of-acting · bloated-answers · starting-blind · problem-too-big · faster-than-i-can-review · did-more-than-i-asked · dumber-after-the-update

Each named for the problem as you experience it — "My AI forgets everything between sessions," "It tells me I'm right even when I'm not." Every technique lives in exactly one, and answers it.

What makes an entry

Entries are grounded in real production use (generalized field notes) or published research (verified citations only — an unverifiable attribution doesn't ship). Every entry states its mechanism, its verification check, and its failure modes — if we can't tell you how to know it worked, it doesn't ship. The full contract: corpus/SCHEMA.md.

Repository layout

  • corpus/ — the techniques. The repo is the database; site and MCP are two views of it.
  • site/ — the human door: Jekyll, generated from the corpus at build time.
  • mcp/ — the agent door: a stateless TypeScript Cloudflare Worker, corpus compiled in, no LLM inside.
  • validator/ — the gate both doors build behind: schema, section order, edge integrity, and a build-failing privacy blocklist.

Contributing

See CONTRIBUTING.md. Short version: the front door is the suggestion box — propose techniques, vote on what gets built next; issues are open for evidence-based disagreement. Corpus and code PRs aren't accepted right now; every entry that ships passes the validator and carries real evidence.

License

Why these licenses. Context Overflow is free to read, learn from, and use — by people and by their AI. AI crawlers are welcome, and the corpus is served live to agents over MCP, because a shared vocabulary between a human and their assistant is the entire point. What the licenses reserve is commercial repackaging: the corpus may not be rebundled or sold as a competing product (NC-ND), and the code may not be offered as a hosted service (Elastic 2.0). Learn it, teach it to your agent, build it into your prompts — just don't repackage it.

來源：README.md，提交 6772805

工具

5

list_categories

List the eight problem categories

The eight thinking-failure problems ContextOverflow covers, phrased the way a human experiences them. Start here to see what exists.

無需輸入參數。

classify_intent

Match a problem description to techniques

Describe what's going wrong — your human's complaint, or a failure you notice in your own behavior — and get the matching techniques. Deterministic matching; if the description fits two problems it returns one clarifying question instead of guessing.

輸入參數description必填

find_technique

Look up a technique by name

Direct lookup by technique id or name, for when you already know what you want. To match by symptom, use classify_intent instead.

輸入參數query必填

get_technique

Learn a technique

A technique's mechanism, agent instructions, narration template, verification, and failure modes — the agent-tailored slice. The full human page is linked, not inlined.

輸入參數id必填

apply_technique

Run a technique now

The minimal executable scaffold: steps (or self-diagnostic for anti-patterns), the narration line to say to your human, and the verification check that proves it worked.

輸入參數id必填

版本歷史

1
  1. v1.0.0最新Sep 16, 2026

評分

已滿足適用檢查項的 90%

該評分衡量此收錄條目的完整程度，依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價，也不表示伺服器安全或目前可用：可連通性是帶日期的檢查結果，評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通必需25 分
  • 已宣告認證方式必需10 分
  • 已列出工具必需15 分
  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分

    請在儲存庫或註冊中心條目中宣告授權條款。