Fabric Spec Mcp

io.github.mrfentmen由 GitHub 帳號 mrfentmen 發布v1.0.0更新於 Nov 13, 2026

Fabric MCP: parse fabric specs, validate composition, weight, width and shrinkage.

發布方自己的服務:Fabric

概覽

AI 產生的概覽

本機 MCP 伺服器,讓助理尋找、解析、驗證、搜尋、比對並寫入 Fabric Spec(.csv)文件。

可以做什麼
  • 在目錄樹中尋找、讀取與寫入 Fabric Spec .csv 檔案,寫入時預設不覆寫既有檔案
  • 將文件解析為記錄、欄位與中繼資料,並把所有記錄匯出為 JSON
  • 依結構規則驗證並給出行號層級的問題,彙總中繼資料與各類別數量
  • 列出記錄 id 與類別,依 id 取得單筆記錄,以純文字或正規表示式搜尋記錄
  • 計算各欄位的數值統計,並比對兩份文件的新增、刪除與變更記錄
  • 回傳精簡的格式參考與一份合法的小型範例文件
適用情境
適合在本機專案中維護 Fabric Spec 文件時使用:讓助理檢查結構錯誤、查詢或搜尋記錄、比較不同版本,或起草新的規格檔案而不必手動編輯 CSV。
執行需求
以本機 stdio 程序執行,從 npm 套件 @mrfentmen/fabric-spec-mcp 安裝;需要 Node.js,並能存取存放規格檔案的目錄。未宣告任何驗證、環境變數或標頭。
安裝前請注意
寫入工具會在磁碟上建立檔案,當 overwrite 設為 true 時會覆寫既有檔案,請確認目標目錄並檢查寫入結果。檔案內容會被助理讀取,避免讓含機密的規格檔案暴露給助理。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Fabric Spec Mcp,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

awesome-servers

One repo for three families of servers, each subtree keeping its own layout and tooling so it can still be built, run and published on its own.

SubtreeWhat it holdsRun one
mcp/338 TypeScript MCP servers, plus 1107 published packages in published/cd mcp/servers/<slug> && npm install && npm run build
acp/79 Python ACP agentspython3 acp/agents/<name>/agent.py
a2a/37 Python A2A servers plus the shared a2a_kitpython3 a2a/servers/<name>/server.py
watchd/a threshold watcher that runs the feeds between editor sessionspython3 watchd/watchd.py --config conditions.json --loop
site/the published site: registry with search, the public-feed atlas, watchd planshttps://mrfentmen.github.io/awesome-servers/

Most servers read public data with no key of their own; the ones wrapping a vendor API take that vendor's key. Nothing in this repo asks a caller to pay. On 2026-10-17 the licence machinery was removed for good: the per-tool src/license.ts gates are gone from the MCP servers, and the A2A and ACP trees no longer carry the hosted licence-gateway half — the skills that used it now answer with an explicit not-implemented stub. Nothing here asks a caller to pay, and nothing they run is refused with a buy link.

Nothing here is generated: each subtree's README says what its servers read, and every server has its own unit tests (mcp/servers/<slug>/tests, acp/agents/<name>/tests, a2a/servers/<name>/tests) plus a live smoke check (tools/smoke*.py, a2a/tools/smoke_all.py).

Novel and original

Nearly all of this is first-of-its-kind work, not a port of an existing catalogue. Before each server or agent was built, it was checked against the official tool lists (the MCP registries, agentclientprotocol.com's agent list and registry) and a live GitHub search for the same domain, and the research logs record what was found and why it still counted as missing — see mcp/MISSING.md, acp/MISSING.md and a2a/docs/. Highlights:

  • The MCP side holds TypeScript servers, one npm package each, and 1039 published packages under published/; mcp/MISSING.md documents the check that preceded every build: what was searched, what was skipped and why the domain still counted as uncovered.
  • On the ACP side there were zero domain agents before this repo — every agent on the official list was a coding agent. These 79 Python agents are the first non-coding ACP agents, and several use protocol surfaces no other agent uses in production (session/request_permission for dataset reads, fs/write_text_file for a briefing file, session/cancel for a conditional watch, image content blocks, a live SSE push feed), each proved-unused in the wild by search at build time. Two more firsts carry their own proof: the orchestrator is an agent whose feeds are agents — it spawns sibling repo agents as real ACP child processes and merges their answers — and the nyc pack is the first multi-agency city surface.
  • acp/MISSING.md doubles as the public-feed atlas: every keyless feed probed live, with its date, what it answered and whether it still works. The published site renders that table at https://mrfentmen.github.io/awesome-servers/atlas.html.
  • The pattern across all three subtrees is the same thesis: wrap one real, keyless public dataset, answer honestly about what the feed did and did not say, and keep the whole thing small, tested and self-contained.

Publishing

  • MCP servers publish from mcp/servers/<slug> (its own package.json, version and README).
  • ACP and A2A servers are repo-run today (Python, no packaging step).

Built on top of it

Creatures that use this corpus as their world live in the sibling repo mrfentmen/superagents (public, with its own claim-check in MISSING.md):

  • polymath — routes your question to the most fitting agents of this repo, spawns each as a real ACP child process, and merges the answers into one brief with per-row citations; a sibling that fails appears as a named row, never silence. Its registry is a data file and its ranking is deterministic word-scoring, not a model opinion.
  • sentry — one plain sentence ("watch air quality in Denver over 150") becomes a tracked watch: a permission-asked protocol write produces the conditions file this repo's watchd/ obeys unchanged, plus the activation lines. Checks it can author today: air thresholds, quake magnitudes within a radius, river gauges at flood stage.
  • fedora hat guy — the persona agent in his own repo mrfentmen/fedora-hat: says "never miss" and means it — one sentence installs the full crown of watches (air + quakes, gauge-named flood) in one permissioned write, and your other questions are forwarded to the siblings above with every row cited.
  • instink — the instinct agent in his own repo mrfentmen/instink: compiles a skill into a reflex — a matcher plus the answer it fires, with hit/miss counters and honest promotion/demotion. The skill doc stays one step away, never deleted.
  • contextink — the context agent in his own repo mrfentmen/contextink: you say "do this" and it acquires the context from the repo, git, and recent files so you don't have to explain it; asks only the minimal high-information-gain questions when acquisition fails.
  • rewind — the replay agent in his own repo mrfentmen/rewind: replays agent runs and mutates one variable to see what would have happened. Replay + mutate + diff for agent traces.
  • scopink — the scope agent in his own repo mrfentmen/scopink: checks if an action is within the scope of what was asked. Detects scope creep in agent plans.
  • curious — the curiosity agent in his own repo mrfentmen/curious: proactively seeks information it doesn't have. Scans for knowledge gaps, seeks answers, learns facts.
  • bored — the idle agent in his own repo mrfentmen/bored: uses idle time productively. Consolidates memories, explores areas, prepares for tasks.

Point the polymath at a checkout of this repo with AWESOME_SERVERS_HOME=$PWD/awesome-servers and it federates across the two repos live.

Published snapshot (published/)

published/ holds one directory per npm package, byte-for-byte what the registry serves: 1039 packages, every version matching npm latest.

  • published/INDEX.md is generated from the tree by python3 scripts/index-published.py.
  • scripts/sync-published-snapshot.py re-downloads every package from npm; run it only when every local fix has been published, because it overwrites dist/ from the tarballs.
  • One package carries a licence gate (MCP_LICENSE_KEY): @mrfentmen/vg-resource-mcp. Another 495 still ship tools that answer Not yet implemented; published/INDEX.md lists every one of them.
  • Every package links to this repo (mrfentmen/awesome-servers) as its source.

servers/ holds the older per-server source that predates the snapshot. Where the two disagree, published/ is what users actually install.

來源:README.md,提交 4c6c31d

工具

14

detect_files

Find Fabric Spec files (.csv) in a directory tree.

輸入參數pathmaxDepth

read_file

Read a Fabric Spec file, optionally a line range.

輸入參數cwdpath必填endLinestartLine

write_file

Write a Fabric Spec file. Refuses to overwrite unless overwrite is true.

輸入參數cwdpath必填content必填overwrite

parse

Parse Fabric Spec into records, fields and document metadata.

輸入參數cwdpathcontent

validate

Validate Fabric Spec against its structural rules and report every issue with a line number.

輸入參數cwdpathcontent

summarize

Summarize a Fabric Spec document: metadata, record counts and per-kind totals.

輸入參數cwdpathcontent

list_records

List Fabric Spec record ids and kinds, with a limit.

輸入參數cwdpathlimitcontent

get_record

Return one Fabric Spec record by id, with all parsed fields.

輸入參數id必填cwdpathcontent

stats

Numeric statistics for Fabric Spec fields across all records.

輸入參數cwdpathfieldcontent

to_json

Export every parsed Fabric Spec record as JSON.

輸入參數cwdpathcontent

diff

Compare two Fabric Spec documents: added, removed and changed records.

輸入參數cwdleftPathrightPathleftContentrightContent

format_reference

Return a compact Fabric Spec reference: structure, required fields and common pitfalls.

無需輸入參數。

sample

Return a small but valid Fabric Spec example document.

無需輸入參數。

版本歷史

1
  1. v1.0.0最新Nov 13, 2026

評分

已滿足適用檢查項的 80%

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
  • 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。

其他 Fabric 整合