CoreScout

com.tensorust由 tensorust.com 發布v1.4.0更新於 Nov 15, 2026

Operational memory for the machine your AI works on: what failed here, what is verified, its state

發布方自己的服務:core.blue

已驗證STDIO僅桌面知識與記憶開發者工具

概覽

AI 產生的概覽

本地優先的機器記憶:觀察指令、失敗與已驗證的修正,讓 AI 在動手前先問這台機器發生過什麼問題。

可以做什麼
  • 查詢這台機器上哪些操作會失敗、哪些修正已被驗證,並附上證據
  • 透過命令列執行並精確量測指令,記錄結束碼、耗時、重試以及結果是否被檢查
  • 檢視已學到的模式、反覆出現的失敗、待驗證的假設及每項結論的依據
  • 連接 AI 用戶端,自動回報其工具呼叫,並設定它可執行的範圍
  • 隨時暫停、查看儲存了哪些資料並刪除紀錄
適用情境
適合 AI 在同一台機器上反覆重新發現相同的建置、部署或測試失敗,並希望這種經驗能跨工作階段保留的情況。適用於單機本地開發,想要有證據支撐的提示而非隨手寫的筆記。
執行需求
需要本地安裝:Windows 安裝程式、macOS 或 Linux 壓縮檔,或以 Rust 安裝命令列、服務與 MCP 橋接程式。不需 Node、Python、Docker 或管理員權限。未宣告帳號、API 金鑰或環境變數;MCP 橋接是註冊到用戶端的本地執行檔。
安裝前請注意
它每秒多次觀察機器,並記錄指令、工具呼叫、結束碼、工作資料夾和工作階段時間,資料保存在本機使用者本地應用程式資料目錄。下載檔尚未簽章,系統可能提示未知發行者,執行前請核對雜湊值。Suggest 以上模式可自行套用變更,建議保持預設或設定限制,並可使用暫停功能。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 CoreScout,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。

Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

CoreScout

Your AI stops rediscovering the same failures every session: CoreScout remembers what happened on your machine and tells your AI before it repeats a mistake.

Why this over the alternatives: Notes files and memory prompts only hold what someone wrote down. CoreScout observes it, and labels what it has verified separately from what it has only seen together. Free, open source, local-first, with an MCP server.

Guarantee: if it doesn't do what this README says, open an issue and I'll fix it.

Your AI forgets everything between sessions. CoreScout lets the computer remember what went wrong, test what actually fixes it, and tell the next AI.

[crates.io] [Release] [License: MIT or Apache-2.0]

[Your AI and your machine feed CoreScout. It sees, learns, verifies with coin-flip trials, and hands improvements back.]

Free, open source, and local-first: no account, no server, no telemetry. Written in Rust, with a desktop app and an MCP server so Claude Code, Claude Desktop, Cursor or any MCP client can ask it questions.

Install

You haveGet
Windows 10 or 11CoreScoutSetup.exe (desktop app + background service + MCP bridge)
Mac (Apple Silicon or Intel) or Linuxthe .tar.gz for your machine from Releases
Rustcargo install corescout-cli corescout-service corescout-mcp

No admin rights, and no Node, Python, WSL or Docker. The downloads are not signed yet, so Windows may say "unknown publisher" (More info, then Run anyway) and a Mac wants right-click, Open. Every file's SHA-256 is in SHA256SUMS.txt on the release page.

Three steps

  1. Install it (above). It starts watching the machine straight away.
  2. Connect your AI. For Claude Code that is one line:
    bash
    claude mcp add corescout --scope user -- "C:\Program Files\CoreScout\corescout-mcp.exe"
    The app's AI screen writes the config for other clients, and can add a hook so every tool call is recorded without the AI having to remember.
  3. Work as usual. Before something risky, your AI can ask "what goes wrong here?" and get an answer that comes with its evidence.

A real answer from a real machine

[CoreScout's real answer on the author's machine: a rebuild chain that failed 5 of 5 times and was never checked afterwards, and a ping that failed 6 of 6.]

Those numbers are what CoreScout returned on the author's PC on 2026-10-04, after a few weeks of real use. It caught a rebuild script that failed every time and was never checked afterwards. It has not yet promoted anything to Verified there, because a claim only gets that label after enough coin-flip trials. It would rather tell you that than invent a win.

Seen together vs. Verified

[Illustration: on the left, two things keep happening together and are labelled Seen together. On the right, a coin flip decides whether to try the fix, and only the measured result is labelled Verified.]

Most tools that "learn" just notice that two things happened together and act on it. That is how you end up with superstitions. CoreScout keeps the two apart, and the labels never blur:

labelmeans
Seen togetherTwo things kept happening together. CoreScout has not tested whether one causes the other.
VerifiedCoreScout let a coin flip decide whether to try the change, then measured the difference.

It also never prints 0% for something it has never measured. A capability that has never run reads never used, because zero out of zero is not zero per cent.

The problem

Your AI starts every session from nothing. The build that fails the same way each time, the deploy that reports success forty seconds before the service is reachable, the test that only breaks when the machine is busy: it rediscovers all of that, every time, and then the window closes.

Nobody writes any of it down, because the thing that would know is the computer, and computers do not learn from their own experience.

The model doesn't learn from every task. Its computer does.

What CoreScout does

  1. Sees. It observes the machine several times a second, and observes what your AI does to it: commands, exit codes, files touched, retries, and (the part that matters most) whether anything actually checked the result.
  2. Learns. It notices what recurs. Which operations fail here, what the runs that worked had in common, which state the machine was in at the time.
  3. Verifies. Then it does the part almost nobody does. A correlation is not a cause, so on a small fraction of occasions CoreScout decides by coin flip rather than by belief, and only those trials support a claim that one thing causes another.
  4. Improves. What survives becomes a capability, with its evidence attached. You approve it. Your AI can then use it, and so can the next AI you connect, because what was learned belongs to the machine.

Everything stays on your computer. No account, no server, no telemetry.

What you will see

After a while, the Home screen says something like:

Your computer is learning how to work with Claude Code.
Today  3 useful patterns learned  1 capability created  2 beliefs retired
Latest  Regenerate schema, then build            Verified  Builds here read a generated file that goes stale.  strong evidence · 94% of 12 uses worked

Open any of it and you get the evidence: what was observed, how many times, whether it was tested or merely noticed, what the alternative was, and what happened. That is the whole point. A tool that tells you what to do without telling you how it knows is asking for trust it has not earned.

How much it may do

Four modes, and you can change them at any time.

ObserveWatches and learns. Changes nothing.
SuggestRecommends improvements. You approve every change. (the default)
AssistMakes small reversible changes on its own. Anything bigger waits for you.
AutopilotApplies changes it has verified, inside limits you set.

Raising the mode never widens what CoreScout may touch; that is a separate setting, and no mode overrides it. There is a Pause that stops everything immediately, and it is checked before anything else in the system. See SECURITY.md.

Command line

corescout run -- <cmd>    Run something, measure it exactly, record itcorescout connect         Set an AI up to report its work automaticallycorescout status          Is it running, and what is connectedcorescout learned         Everything it has learnedcorescout explain <id>    The evidence behind one of thosecorescout failures        Operations that recur and go wrong herecorescout hypotheses      What it is testing and cannot yet answercorescout ai setup        How to connect an AIcorescout mode assist     Change how much it may docorescout pause           Stop everything nowcorescout privacy         Exactly what is stored, and where

--json on any of them gives the same object the application receives.

What is actually here

text
your machine     ↓computational mirror        what this computer physically is, right now     ↓machine experience          what it has been like to be this machine     ↓science                     falsifiable claims, and what refutes them     ↓concepts                    structure it found and named itself     ↓AI operational experience   what working with an AI has taught it     ↓capabilities                procedures that survived being tested     ↓your AI                     through one standard tool interface     ↓bounded changes             scoped, reversible, audited, visible     ↓your machine again
  • PRODUCT.md: what each screen is for and why it is worded that way
  • DEMO.md: a real trap, built for real, with and without an experienced computer
  • MCP.md: the tools your AI gets, and how to connect one
  • CAPABILITIES.md: how a correlation becomes something runnable
  • PRIVACY.md: everything that is stored, and where
  • SECURITY.md: the boundaries, and how they are enforced
  • WINDOWS.md: building, packaging, signing
  • STORE.md: the Microsoft Store package, and what a submission needs
  • ARCHITECTURE_PRODUCT.md: the product layers and the rules they follow
  • ARCHITECTURE.md: the research crate graph and what each layer may not do

The first run on real hardware

The first time this ran on hardware (a 13th Gen Intel Core i7-13700KF, 16 physical cores, 24 logical), it made over 900 reflections in 36 seconds:

text
86 observable parts × 9 channels32 recurring states discovered, which nothing told it to look for3 concepts promoted, 29 refusedknowing which state it was in reduced prediction error by 42.4%

Its own account of itself, generated from what it had measured:

I distinguish 32 recurring states of myself, which I found rather than being told about.

I do not predict my own next state better than assuming nothing changes.

Both of those are in REAL.md, along with the lineage experiment that returned −4.1% and the four bugs only real hardware could expose. The project's habit is to publish the results that go the wrong way, and the product kept it.

The research underneath

CoreScout began as an experiment about whether a machine can discover useful structure in itself: falsifiable hypotheses, concepts it coins and refuses, association held apart from causal evidence, beliefs withdrawn when the evidence ages out. All of that is still here and still runs, under corescout-lab.

Building it

pwsh scripts/release.ps1 builds CoreScoutSetup.exe into dist/ (details in WINDOWS.md); cargo build --release -p corescout-cli -p corescout-service -p corescout-mcp builds just the programs.

bash
cargo test --workspace       # ~1,300 testscargo clippy --workspace --all-targetscargo fmt --all -- --checkpwsh scripts/release.ps1     # installers into dist/

The desktop shell is deliberately outside the Cargo workspace, so the tests above still run in seconds.

Privacy Policy

CoreScout is local-first. Everything it records stays in %LOCALAPPDATA%\CoreScout\ on the machine that produced it.

  • Collected. Processor layout and live performance counters; the commands and tool calls AI clients make, with credentials stripped as the data enters; exit codes, durations, retries, and whether anything verified the result; the folder each piece of work happened in; when each AI session ran.
  • Never collected. The contents of your source files, your prompts, model responses, and any token, key or password.
  • Used for. Forming and testing theories about how this machine behaves, so the AI tools connected to it can be told what has failed here before and what has been verified. Nothing else.
  • Shared with third parties. Nothing. There is no network endpoint to share it with: no crate in this project opens an outbound connection.
  • Retention. Kept until you delete it. The application deletes any category of it on request, and removing %LOCALAPPDATA%\CoreScout\ removes all of it; CoreScout then starts again from nothing.
  • Contact. https://github.com/Mattbusel/corescout/issues

Full policy: docs/PRIVACY.md

Licence

MIT or Apache-2.0, at your option.

來源:README.md,提交 55ac60d

工具

0
SourceWeft 曾在隔離沙箱中以無憑證方式啟動該軟體包一次,以列出其工具,但未能完成:該軟體包類型無法在沙箱中執行。

版本歷史

1
  1. v1.4.0最新Nov 15, 2026

評分

已滿足適用檢查項的 50%

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25 分

    請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。

  • 已宣告認證方式10 分

    請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。

  • 已列出工具15 分

    請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分

其他 Core 整合