Asynthetic

io.github.asyntheticai由 GitHub 帳號 asyntheticai 發布v0.1.3更新於 Oct 24, 2026

Ground truth for AI coding agents: what breaks between library versions and how to fix it

已驗證Streamable HTTP可網頁執行知識與記憶開發者工具

概覽

AI 產生的概覽

Asynthetic 為程式助理提供人工整理的遷移圖譜:函式庫版本之間的破壞性變更,附修改前後程式碼與引用來源。

可以做什麼
查詢某個套件在兩個版本之間的破壞性變更,每筆含修改前後程式碼與引用來源,並附驗證等級與最後核驗日期。查詢某個版本會破壞什麼。檢查兩個套件版本是否相容,回傳 true、false 或 unknown 並附引用理由。列出伺服器目前收錄的遷移圖譜。沒有資料時回傳 found: false,並提示助理不要猜測。
適用情境
當助理要升級 JavaScript 或 AI SDK 類別函式庫(例如 Next.js 14 升到 15、AI SDK 4 升到 5),需要真正的差異而非現行文件時使用;也可在改程式碼前檢查 peer 版本相容性。
執行需求
可使用 Streamable HTTP 的遠端託管端點,或以 npx 透過 stdio 執行的本機 npm 套件,後者離線提供內建圖譜。正常使用不需帳號、API 金鑰或環境變數。自行託管 HTTP 模式使用 PORT,從 Postgres 提供資料使用 SUPABASE_URL 和 SUPABASE_ANON_KEY。
安裝前請注意
所有工具都是對整理資料的唯讀查詢:不執行程式碼、不在請求時抓取網路、不存取你的專案或檔案。託管端點只收到工具參數(套件名稱、版本字串),工作階段為暫存於記憶體且彼此隔離。涵蓋範圍較窄(目前為 Next.js 14 到 15 和 AI SDK 4 到 5),且處於免費公測階段。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Asynthetic,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/io.github.asyntheticai%2Fasynthetic/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": {
    "asynthetic": {
      "type": "http",
      "url": "https://asynthetic.up.railway.app/mcp"
    }
  }
}

README

Asynthetic

Ground truth for AI coding agents.

Your agent's knowledge froze on a date. Asynthetic tells it exactly what changed — every breaking change between library versions, cited and tested.

sh
claude mcp add --transport http asynthetic https://asynthetic.up.railway.app/mcp
jsonc
// agent asks: get_migration("ai", "^4.0.0", "^5.0.0"){  "found": true,  "verification_level": "medium",  "migration": {    "breaking_changes": [      {        "title": "maxTokens renamed to maxOutputTokens",        "before_code": "generateText({ model, maxTokens: 1024, prompt })",        "after_code":  "generateText({ model, maxOutputTokens: 1024, prompt })",        "source_url": "https://ai-sdk.dev/docs/migration-guides/migration-guide-5-0"      }      // …22 more, every one cited    ]  }}

Models improve every month. Their knowledge still ends on a date.

Every coding agent ships with a knowledge freeze date. The libraries it writes against don't.

So agents do what the training data taught them: they write maxTokens into AI SDK 5 (removed), synchronous params into Next.js 15 (now a Promise), LangChainAdapter imports from a package that no longer exports them. The code looks right. The agent is confident. The build breaks — or worse, it doesn't, and the bug ships.

Tools that fetch current documentation tell agents what the API looks like today. That's half the problem. An upgrade needs the diff: what changed, what replaced it, and what silently behaves differently. No amount of "current docs" contains that.

Asynthetic is that missing half. Ask it about an upgrade, and it returns a hand-curated migration map: ordered breaking changes with before/after code, deprecation timelines, peer-version requirements, and a citation for every claim. Retrieval, never generation — the same question gets the same verified answer every time.

Before / after

Without Asynthetic — the agent "upgrades" your AI SDK app from memory: keeps result.usage for billing (now silently counts only the final step), leaves temperature unset (v5 removed the implicit 0 — output turns nondeterministic), renames nothing because v4 syntax was correct when the model trained.

With Asynthetic — one tool call returns all 23 breaking changes for that exact version window, including the three above, each with the fix and the official source. The agent applies facts. You review a diff instead of debugging a regression.

That's the product: the difference between an agent that remembers and an agent that knows.

What a map gives your agent

Migration MapsHand-curated upgrade intelligence per version window: 40 breaking changes across 2 live maps (next 14→15, ai 4→5)
CitationsEvery map cites official sources and carries a last_verified date. Maps are never LLM-generated.
Verification LevelsEvery response self-reports confidence: high (empirically tested on real builds), medium, low — derived, never self-declared
Field NotesWhere reality diverges from docs, maps say so — e.g. Turbopack doesn't actually enforce the @next/font removal that Webpack does; Next 15's React 19 "requirement" isn't build-enforced. Found by running the upgrades, not reading about them.
SemVer ResolutionAsk with ranges (^14.2.0 → ^15.0.0); resolution method is always disclosed
Peer ChecksStatic compatibility answers (next@15 + [email protected] → false, with the cited reason)
Honest MissesNo data means found: false and an explicit instruction not to guess — never a plausible fabrication

The next 14→15 map is Asynthetic's first verification_level: high: every entry tested against real builds on 14.2.35 and 15.5.20, across Turbopack and Webpack, dev and production.

Install

Hosted (nothing to install):

sh
claude mcp add --transport http asynthetic https://asynthetic.up.railway.app/mcp

Local (offline, bundled data):

sh
claude mcp add asynthetic -- npx -y asynthetic

Any MCP client works — Cursor, Claude Code, Windsurf, or your own agent:

json
{ "mcpServers": { "asynthetic": { "url": "https://asynthetic.up.railway.app/mcp" } } }

Then upgrade something. The agent calls get_migration before it edits.

Tools

ToolAnswers
get_migration(package, from, to)"What breaks between these versions, and how do I fix each one?"
get_breaking_changes(package, version)"What does this version break?"
check_peer_compatibility(a, va, b, vb)"Do these two versions work together?" — true / false / unknown, cited
list_available_maps()"What does Asynthetic know?"

Full response shape, transports (Streamable HTTP + stdio + legacy SSE), and self-hosting: see docs below.

Why trust it

Asynthetic's founding rule: a wrong map is worse than no map. Everything follows from that.

  • Curated exclusively from official changelogs, migration guides, and releases — with the URL attached to every entry.
  • verified status is earned empirically, not editorially: real projects, both bundlers, dev and production, on pinned versions. The process regularly finds things the docs get wrong — those findings ship in the maps.
  • Confidence is machine-readable (verification_level, source_count, pre-release warnings), so agents can weigh answers instead of trusting blindly.
  • Unknown is a first-class answer. The server tells agents when it doesn't know and instructs them not to invent.

Narrow and deep, on purpose: fast-moving AI and JavaScript frameworks first, correctness over coverage always.

Self-hosting

sh
git clone https://github.com/asyntheticai/asynthetic.gitcd asynthetic && npm install && npm run build && npm start

Zero-config serves the bundled maps over stdio. Set PORT for HTTP mode (/mcp, legacy /sse, health at /). Set SUPABASE_URL + SUPABASE_ANON_KEY to serve from Postgres (schema/schema.sql, then npm run seed). Missing config never crashes — it falls back and says so on stderr.

data/maps/          the maps (source of truth, JSON, Zod-validated)src/server.ts       MCP toolssrc/store/          Postgres + local backends, SemVer resolverscripts/smoke.ts    23-check end-to-end suite across every transport

Contributing a map

  1. Official sources only. Changelogs, release notes, migration guides. Record every URL. Blog posts can start a hunt; they can't finish one.
  2. Follow the schema in src/types/migration-map.ts — validation rejects anything malformed before it can ever be served.
  3. New maps enter as draft. verified is earned by testing snippets on real builds of the pinned versions. Retired maps become stale — excluded from serving, kept as record.
  4. npm run smoke must pass. Don't guess versions or API names: if it's not in an official source, flag it in the PR instead.

Status

Free public beta — hosted endpoint and npm package open to everyone. Next: more maps (Tailwind 3→4, React 18→19, Zod 3→4, AI SDK 5→6), more high-verification promotions, project-aware compatibility analysis. Details will be announced.

Security

Read-only by construction: every tool is a lookup over curated data. No code execution, no request-time fetching, no access to your project or filesystem. The hosted endpoint sees only tool arguments (package names, version strings); sessions are ephemeral, in-memory, and isolated. Report issues via GitHub security advisories.

License

Source-available under BSL 1.1: self-hosting for your own use — personal, team, company-internal — is always permitted; reselling it as a competing hosted service isn't. Converts automatically to Apache 2.0 on July 4, 2030.


Built with TypeScript, the official MCP SDK, Zod, and Supabase. Serving live at asynthetic.up.railway.app/mcp.

來源:README.md,提交 844feff

工具

0
SourceWeft 尚未讀取此伺服器的工具。

版本歷史

1
  1. v0.1.3最新Sep 16, 2026

評分

已滿足適用檢查項的 75%

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

    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

    請在儲存庫或註冊中心條目中宣告授權條款。