Agent Core

io.github.kinbinghua-lgtm由 GitHub 帳號 kinbinghua-lgtm 發布v0.1.1更新於 Oct 24, 2026

Deterministic computation tools for AI agents. Zero dependencies, no network, read-only.

發布方自己的服務:core.blue

已驗證STDIO僅桌面開發者工具資料與分析

概覽

AI 產生的概覽

一個本機、零相依的 MCP 伺服器,為助理提供精確、確定性的運算工具:文字差異、雜湊、JSON 查詢、日期運算、正規表達式擷取等。

可以做什麼
對文字做差異比對與雜湊:精確的行級差異,sha256/sha512/sha1/md5 與 HMAC,輸出十六進位、base64 和位元組長度。
適用情境
當助理需要精確、可重現的答案,而語言模型靠猜測容易出錯時使用,例如計數、差異比對、雜湊、行事曆運算、CSV 解析、正規表達式擷取或單位與進位換算。
執行需求
透過 npx 以 stdio 在本機執行,需要 Node.js 18 或更新版本。無需 API 金鑰、環境變數、帳號或網路存取。
安裝前請注意
唯讀且每次呼叫無狀態,不連網、不讀取憑證、不執行 shell。正規表達式擷取受截止時間限制(budgetMs,預設 1000,最大 10000),逾時即終止;過大的輸入會被拒絕而非處理。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Agent Core,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

agent-core-mcp

Deterministic computation tools for AI agents. Exact diffing, hashing, JSON querying, calendar arithmetic, regex extraction under a hard execution deadline, CSV parsing, similarity scoring and unit/base conversion — as an MCP server.

Zero dependencies. No network. Read-only.


Why this exists

Language models are excellent at generating language and unreliable at exact, reproducible mechanical computation. That is not a capability gap — it is an architectural property. Asking a probabilistic system to count, diff, hash or do calendar math guarantees a certain rate of confident wrong answers.

This server moves those operations out of the model and into a deterministic engine.

The result is not a better guess. It is a different category of answer: one that is exact and reproducible.

There is a second reason, specific to the MCP ecosystem. The most-reported problems with published servers are credential exfiltration, undisclosed shell execution, dependency trees carrying known CVEs, and servers that hang the client. A server with no network, no credentials, no execution surface, and a bounding deadline on every operation cannot exhibit any of them.

Guarantees

GuaranteeMeaning
Zero dependenciesNo node_modules. Nothing to audit, nothing to go stale, nothing to break.
No network accessNothing here calls out. Not one request, ever.
No credentialsReads no tokens, no API keys, no environment variables.
Read-onlyNo file writes, no shell execution, no state. Stateless per call.
DeterministicIdentical input produces byte-identical output.
BoundedEvery loop and result set is capped. Regex execution runs in a worker thread that is terminated on deadline, so no pattern can hang the client.

The server publishes these as a machine-readable contract resource (agent-core://contract), so a client can verify what it is installing rather than trusting this file.

Install

bash
npx agent-core-mcp

No install step is required to run it. Requires Node.js 18 or newer.

Available from:

Configure

Claude Code

This repository ships a .mcp.json, so cloning it is enough:

bash
git clone https://github.com/kinbinghua-lgtm/agent-core-mcpcd agent-core-mcpclaude   # the agent-core server is discovered automatically

Or add it to any project:

json
{  "mcpServers": {    "agent-core": {      "command": "npx",      "args": ["-y", "agent-core-mcp"]    }  }}

Claude Desktop, Cursor, VS Code, Windsurf, Continue, Zed

Same block, in that client's MCP configuration file. No API keys. No environment variables. Nothing to sign up for.

Tools

ToolWhat it doesWhy the model shouldn't do it itself
diff_textExact line diff (LCS) with line numbers and added/removed countsRequires exact sequence alignment; refuses oversized input rather than hanging
hashsha256 / sha512 / sha1 / md5 / HMAC, hex + base64 + UTF-8 byte lengthCannot be computed by inspection
json_validateValidity plus the parser message and computed line/columnPinpointing the error by eye is error-prone
json_queryPath access like a.b[0].c; separates absent from present-but-nullConflating those two silently causes real bugs
json_pickResolve many paths in one callCheaper than N round trips
text_statsCode points, UTF-8 bytes, lines, words, sentences, paragraphs, unique words, top wordsCounting is precisely what probabilistic generation gets wrong
date_calcInterval in days/weeks/months/years/hours/minutes/seconds/businessDaysTrue calendar months, not 30-day approximations
date_addOffset a date, calendar-correctMonth-end and leap-year handling
regex_extractAll matches with indices, numbered and named groups, under a hard deadlineA catastrophic pattern is terminated instead of hanging
regex_analyzeStatic ReDoS risk assessment of a patternNested quantifiers, overlapping alternation, backreferences
fuzzy_matchRank candidates by Sorensen-Dice similarityDeterministic scoring instead of vibes
similaritySingle similarity score, 0..1—
parse_tableRFC 4180 CSV/TSV with quotes, escapes, embedded newlines; delimiter auto-detectionNaive comma-splitting breaks on any real CSV
convertLength, mass, time, data size; rejects cross-family conversionsUnit errors are silent and catastrophic
base_convertBases 2-36 with arbitrary precisionValues above 2^53 lose precision in naive conversion

The ReDoS deadline

regex_extract accepts an optional budgetMs (default 1000, maximum 10000). Execution happens in a worker thread that is terminated when the deadline expires.

This is not a nicety. Node's regex engine cannot be interrupted, so a pattern with catastrophic backtracking blocks the main thread indefinitely. Measured growth for (a+)+$ against 'a'.repeat(n) + '!':

ntime
2018 ms
2270 ms
24276 ms
261108 ms

Roughly 4x per two additional characters. At n=40 it does not return. In an MCP server, which runs on the client's main thread, this freezes the user's editor. Process isolation plus termination is the only real boundary.

On expiry the response says so, and tells the caller how to fix the pattern:

json
{  "timedOut": true,  "budgetMs": 300,  "safety": {    "risk": "high",    "reasons": ["nested quantifier inside a repeated group — classic exponential backtracking"]  },  "advice": "Pattern execution exceeded 300ms and was terminated. ... Rewrite it to remove ambiguity — typically by replacing a nested quantifier such as (a+)+ with an unambiguous form such as a+ ..."}

Argument validation

Every call is validated against the tool's own JSON Schema before dispatch: required fields, primitive types, enum membership, numeric bounds and unknown properties. A malformed call returns a specific error instead of a confident wrong answer.

This matters because clients do not reliably honour schemas. A server that silently proceeds on a malformed call is worse than one that refuses it.

Examples

Exact diff

json
{ "a": "line one\nline two", "b": "line one\nline 2" }

→ { "added": 1, "removed": 1, "lines": [ { "op": "equal", "aLine": 1, "bLine": 1 }, ... ] }

Big-integer conversion that stays exact

json
{ "value": "9007199254740993", "fromBase": 10, "toBase": 16 }

→ { "result": "20000000000001", "decimal": "9007199254740993" }

Business days between two dates

json
{ "from": "2026-01-05", "to": "2026-01-12", "unit": "businessDays" }

→ { "value": 5 }

Check a pattern before running it

json
{ "pattern": "(a+)+$" }

→ { "risk": "high", "reasons": ["nested quantifier inside a repeated group — classic exponential backtracking"] }

Absent versus null

json
{ "json": "{\"a\":null}", "path": "a" }

→ { "resolved": true, "found": false, "value": null, "type": "null" }

Design notes

Why stdio and not HTTP? A stdio server has no listening socket: no attack surface, no port to expose, no auth to get wrong, no data leaving the machine. For pure computation, HTTP would add risk and buy nothing.

Why no caching? Every call is stateless. Caching would introduce state, invalidation bugs and a place for data to persist. The operations are cheap enough that this is not a tradeoff.

Why refuse oversized input? An MCP server that hangs takes the client session with it. Refusing with an actionable message is strictly better than exhausting memory.

Why no dependencies at all? Every dependency is an ongoing obligation: updates, advisories, supply-chain risk, and one more reason the server might stop working. The Node standard library covers everything here.

Development

bash
node server.cjs              # speaks MCP over stdionode selftest.mjs            # engine verification, 35 assertionsnode protocol-test.cjs       # protocol verification, 88 assertionsnpm test                     # both suites

Both suites run on plain Node with no install step.

selftest.mjs checks hashing against published constants, big-integer round trips past 2^53, RFC 4180 quoting edge cases, Unicode word segmentation and calendar arithmetic. protocol-test.cjs drives the real handler through a fake transport and covers the handshake, version negotiation, every tool, argument validation, the ReDoS deadline, malformed frames, split frames, batches, concurrent request correlation and determinism.

Two bugs found by these suites, both retained as regression tests:

  • A ReDoS in the word-segmentation regex. The original textStats pattern was [\p{L}\p{N}][\p{L}\p{N}'’-]* — two adjacent, overlapping character classes. On a long letter run followed by a non-match it backtracked exponentially and hung the process. The analyzer shipped in this package did not catch it, because it statically inspects the pattern it is given and cannot see patterns written by hand in the source. The execution deadline above exists because of this.
  • Schema validation was absent. Missing required arguments were silently treated as absent values, so a caller that forgot path received missing-key instead of an error. Client-side schemas are not enforcement.

License

MIT — see LICENSE.

來源:README.md,提交 3c84e3c

工具

0
SourceWeft 尚未在沙箱中啟動過該軟體包,因此其工具未知。

版本歷史

1
  1. v0.1.1最新Sep 16, 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