City Atlas MCP 城市圖譜
Taiwan city data for Claude · 給 AI 用的台灣城市資料
City Atlas aligns Taiwan government open data and OpenStreetMap — traffic crashes, bus service, schools, flood risk, population and more — onto one 125 m grid covering 18 districts of Greater Hsinchu: 82 datasets in 11 categories, down to 161 indicators (141 on the grid plus 20 at neighbourhood / district / county level), plus 47 elections. This public MCP server leaves out 2 datasets (4 indicators) that are derived from Google Street View, which Google's terms do not let us republish, so here you can query 80 datasets and 157 indicators. It lets Claude query that grid and returns every number with its source, data date and caveats. The numbers are computed by the same code as the City Atlas website, not generated by the model.
城市圖譜把政府開放資料與 OpenStreetMap 資料疊到同一張 125 公尺的格網(大新竹 18 個鄉鎮市區):11 大類 82 項資料, 細到 161 條指標(141 條格網 + 20 條里 / 鄉鎮 / 縣市),加上 47 屆選舉。公開的 MCP 少了 2 項 4 條 Google 街景判讀的資料(Google 平台條款不讓我們轉發), 所以這裡查得到 80 項、157 條指標。接上這個 MCP,Claude 就能直接查,回答都附出處、資料日與但書;數字是跟網站同一套程式算的,不是模型猜的。
Connect · 怎麼接
Server URL · 網址:
https://mcp.cityatlas.tw/mcp
- Claude (web or desktop) → Customize → Connectors → Add custom connector.
claude.ai 網頁版或桌面版 → 左邊 Customize → Connectors → Add custom connector。 - Paste the URL above. Authentication: none.
貼上面的網址,不用登入。 - In a chat, open + → Connectors and turn on City Atlas.
對話左下 + → Connectors → 打開城市圖譜。
Claude Code: claude mcp add --transport http cityatlas https://mcp.cityatlas.tw/mcp
Other AIs · 在其他 AI 裡用
Same URL, no login, read-only. Any AI that can add a remote MCP server (Streamable HTTP) can use it. Claude draws the interactive map inside the chat (tested); ChatGPT, VS Code and Cursor say they support the same map view (not tested by us).
Where the chat cannot draw it, the answer carries a link that opens the same map in a browser (phones too).
同一個網址、不用登入、只能讀。只要那個 AI 能加「遠端 MCP 伺服器」就能用。Claude 會把互動地圖直接畫在對話裡(實測過);ChatGPT、VS Code、Cursor 官方說支援同樣的地圖畫面(我們沒測)。
畫不出來的 AI,回答會附一個網址,點開就是同一張地圖(手機也行)。
| AI | How to add · 怎麼加 | Status · 狀態 |
|---|---|---|
| ChatGPT web 網頁版 |
chatgpt.com → Plugins → + → Add custom MCP server → paste the URL, authentication No authentication. 貼上面的網址,驗證選 No authentication。 Per OpenAI: Pro, Business, Enterprise, Edu; Free / Plus not confirmed. The mobile app cannot add custom servers (a server added on the web shows up there). 官方寫 Pro、Business、Enterprise、Edu 可加;Free、Plus 沒寫。手機 App 不能自己加。 |
Not tested yet. From OpenAI's docs, whose menus change often. 沒實測,照官方文件寫,選單常改。 |
| Codex CLI · IDE · app |
codex mcp add cityatlas --url https://mcp.cityatlas.tw/mcp |
Tested 2026-10-11 (CLI): it called the catalog, rank and describe tools. 實測過(CLI)。 |
| Claude Code | claude mcp add --transport http cityatlas https://mcp.cityatlas.tw/mcp | Tested 實測過 |
| Cursor | Add to Cursor · 一鍵加入, or paste into ~/.cursor/mcp.json (below). 或貼下面的設定。 | One-click link not clicked yet; the JSON works the same. 一鍵連結還沒實點,貼設定一樣能用。 |
| VS Code + GitHub Copilot | Add to VS Code · 一鍵加入, or paste into .vscode/mcp.json (below). Company accounts: an admin must allow MCP. 公司帳號要管理員開 MCP。 | One-click link not clicked yet; the JSON works the same. 一鍵連結還沒實點,貼設定一樣能用。 |
| Mistral Le Chat | Connectors → Add Connector → Custom MCP Connector → paste the URL. | Not tested 沒實測 |
| Perplexity | Settings → Connectors → add custom → Remote → authentication None (Pro and above 官方寫 Pro 以上). | Not tested 沒實測 |
| Gemini | Gemini app custom connectors are limited to paid, English-language users in the US, so they are not available in Taiwan yet. The personal Gemini CLI was retired in favour of Antigravity CLI. Gemini App 的自訂連接器目前限美國英文付費用戶,台灣一般使用者還用不到;個人版 Gemini CLI 已停,改 Antigravity。 | Not tested 沒實測 |
Config snippets · 設定片段 (field names as in each vendor's docs · 欄位名照各家官方)
VS Code .vscode/mcp.json
{ "servers": { "cityatlas": { "type": "http", "url": "https://mcp.cityatlas.tw/mcp" } } }
Cursor ~/.cursor/mcp.json
{ "mcpServers": { "cityatlas": { "url": "https://mcp.cityatlas.tw/mcp" } } }
Antigravity CLI ~/.gemini/config/mcp_config.json · Gemini CLI (company accounts, 公司付費帳號) uses httpUrl instead of serverUrl
{ "mcpServers": { "cityatlas": { "serverUrl": "https://mcp.cityatlas.tw/mcp" } } }
Snippets for VS Code, Cursor and Antigravity follow the vendors' docs and were not run by us. 這三段照官方文件組,我們沒實際跑過。
Try asking · 範例問句
- 竹北哪個里,行人事故每千人最多? — Which neighbourhood in Zhubei has the most pedestrian crashes per 1,000 residents?
- 65 歲以上占比最高的里是哪些? — Which neighbourhoods have the highest share of residents aged 65+?
- 2020 到 2024 年,哪些里的總統投票率掉最多? — Where did presidential-election turnout fall most between 2020 and 2024?
- 竹北市光明六路 10 號附近,公車一天幾班?在大新竹排第幾?
- 需要度前 10 名的里是哪些? — Where would a new bus route help the most people?
- 事故多、人行道又少的里有哪些?
- 離公車站超過 500 公尺、又住很多人的格子在哪裡?
- 把竹北的行人事故(每千居民)畫成地圖。 — Map pedestrian crashes per 1,000 residents in Zhubei.
Map preview without Claude · 不用 Claude 先看地圖:https://mcp.cityatlas.tw/map?k=crash_ped4…
Tools · 工具(all read-only · 全部只能讀)
| Tool | 中文 | What it does |
|---|---|---|
search_catalog | 查資料目錄 | Find indicators by keyword (crashes, buses, sidewalks, flooding, trees…; neighbourhood-level age, income and 47 elections, referendums and recalls) |
describe_metric | 看資料說明 | Full provenance of one indicator: agency, dataset, URL, date, coverage, caveats |
rank | 排名 | Rank neighbourhoods (里) or districts by an indicator, optionally per 1,000 residents or per km of road |
lookup_place | 查某個地方 | Numbers for an address, neighbourhood or district, and where it ranks in Greater Hsinchu |
need_ranking | 需要度 | Where most people lose time on public transport compared with driving |
overlay_metrics | 兩樣疊 | Which places are high on two indicators at once (3×3 classes) |
compare_groups | 分組比 | Cells with vs without B, compared within 25 strata of residents × road length |
filter_cells | 篩條件 | 125 m cells that meet up to four conditions, and how many people live there |
compare_change | 看變化 | Which neighbourhoods grew or shrank most between two years: population, crashes, income, election turnout (same as the website's change view) |
show_map | 畫地圖 | Interactive choropleth map inside the conversation (MCP App): drag, zoom, hover for values; always returns a link that opens the map in a browser |
Data · 資料
- Coverage: 18 districts of Greater Hsinchu (Hsinchu County, Hsinchu City and neighbours) on a 125 m grid, aggregated to 里 and 鄉鎮.
- Sources: Taiwan central and local government open data (Open Government Data License, version 1.0), OpenStreetMap (© OpenStreetMap contributors, ODbL), Overture Maps. Each tool result names the agency and dataset.
- Neighbourhood-level only (里, one number per 里, not spread onto the grid): registered population 2018–2025, residents aged 65+ and 14 and under (end of 2025), median comprehensive income from tax filings (Ministry of Finance), and Central Election Commission results — 47 elections, referendums and recalls, turnout and vote shares, as published, with no party colours or labels. These are household-registration and polling records, not where people actually live or any prediction.
- Estimated indicators are marked as estimates. Districts without data are reported as “not collected”, never as zero.
- Rankings mean “worth looking at first”, not “most dangerous”. Results describe; they do not establish causes.
Contact · 聯絡
hello@cityatlas.tw · cityatlas.tw · Privacy · 隱私