bhoonidhi-mcp
bhoonidhi-mcp
一个 MCP 服务器,让 AI 代理能够以自然语言从 ISRO 的 Bhoonidhi 浏览与订购门户(NRSC)搜索、保存、下载和加入购物车卫星影像。代理可以将诸如“去年一月在 Shillong 上空的 Sentinel-2”这样的句子转化为对实时门户的真实搜索,查看实际可下载的内容,保存搜索以便日后复用,预览下载会获取什么,并且——在登录后——下载开放获取的影像或将其加入购物车。
它是 bhoonidhi-downloader SDK 的一个轻量封装——与 bhd CLI 使用的是同一个客户端——因此不会重复实现任何门户逻辑。
状态
搜索和保存无需登录;下载和加入购物车需要登录。 这些工具可以访问 41 颗卫星任务和 79 个传感器的完整存档,搜索实时门户,并将搜索保存为可复用的 slug——所有这些都不需要凭据。下载开放获取影像以及将影像加入 Bhoonidhi 购物车需要登录,可通过 bhd auth login 在带外完成(服务器复用该会话)。
Related MCP server: Google Earth Engine MCP Server
工具
工具 | 功能 | 需要登录 |
| 门户支持的卫星、传感器和搜索词词汇表,实时获取自 Bhoonidhi。 | 否 |
| 将地名(如“Loktak Lake”)转换为质心和边界框。拒绝非地名的输入。 | 否 |
| 基于区域和日期范围的自然语言场景搜索。将非正式的卫星名称解析为精确的搜索词,并报告每个场景的可用性(Ready / Archived / OnOrder / Priced)。无状态——不保存任何内容。 | 否 |
| 预演:显示下载这些结果会获取什么,以及在真正下载之前会跳过什么。 | 否 |
| 将搜索(参数与 | 否 |
| 以简洁摘要列出已保存的搜索:slug、名称、日期范围、卫星、区域和可用性。 | 否 |
| 按 slug 返回一个已保存的搜索及其场景。 | 否 |
| 按 slug 删除一个已保存的搜索。 | 否 |
| 报告是否已配置登录。绝不处理密码或令牌。 | 否 |
| 在后台下载一个已保存搜索的开放获取场景,保存到服务器配置的固定根目录下。立即返回一个 | 是 |
| 按 | 否 |
| 阻塞直到下载完成(或达到超时上限),然后报告结果——这是后台监视器循环所依赖的高效原语。 | 否 |
| 将一个已保存搜索的场景加入购物车(将每个场景路由到 ready / on-order / priced)。 | 是 |
| 列出当前在购物车中的场景。 | 是 |
| 从购物车中移除场景。 | 是 |
可用性很重要:一个 OpenData 场景并不一定已经准备好下载。search_scenes 和 preview_download 会区分 Ready(现在就可以获取)和 Archived(开放数据,但可能需要在门户上先提交请求),这样代理就不会过度承诺。
下载在后台运行,与对话无关:download_query 立即返回一个 job_id,传输过程自行继续。用 download_status 检查一次进度——它会报告已下载字节数、传输速率,以及已知大小时的百分比——或者用 download_wait 跟随任务直到完成(或达到超时上限),这样代理可以委托一个后台监视器并保持对话流畅,而不是空转循环。任务的生命周期与服务器进程相同,因此一旦下载规模较大,状态会建议运行一个独立的 bhd query download <slug> 命令,由你自己掌控。
安装
服务器是一个带有控制台入口点的 Python 包,bhoonidhi-mcp。
使用 uv 从源码安装:
git clone https://github.com/geovicco-dev/bhoonidhi-mcp
cd bhoonidhi-mcp
uv sync这会在 .venv/bin/bhoonidhi-mcp 暴露 bhoonidhi-mcp 命令。服务器通过 stdio 通信,由 MCP 客户端启动——你只需将客户端指向该命令即可。
连接客户端
每个 MCP 客户端都需要同样的东西:要启动的命令。请使用入口点的绝对路径(在 GUI 和 CLI 客户端中最可靠):
/path/to/bhoonidhi-mcp/.venv/bin/bhoonidhi-mcpClaude Desktop / Claude Code
claude_desktop_config.json(或 claude mcp add):
{
"mcpServers": {
"bhoonidhi": {
"command": "/path/to/bhoonidhi-mcp/.venv/bin/bhoonidhi-mcp",
"args": []
}
}
}OpenCode
~/.config/opencode/opencode.json:
{
"mcp": {
"bhoonidhi": {
"type": "local",
"command": ["/path/to/bhoonidhi-mcp/.venv/bin/bhoonidhi-mcp"],
"enabled": true
}
}
}MCP Inspector(无需代理即可试用)
npx @modelcontextprotocol/inspector /path/to/bhoonidhi-mcp/.venv/bin/bhoonidhi-mcp示例
用自然语言询问代理:
“2024 年 1 月 Shillong 上空有哪些 Sentinel-2 场景,其中有多少我可以实际下载?”
代理会先为 Shillong 调用 resolve_location,然后调用 search_scenes,并根据结果回答——例如,所有场景都是 Archived(开放数据,但每个场景可能需要在门户上先提交请求才能下载),而不是声称它们都已就绪。
可尝试的提示词
将这些复制到任何已连接的代理中,感受一下它的能力。
探索存档
“Bhoonidhi 有哪些卫星和传感器?”
“ResourceSat-2A 搭载了哪些传感器,分辨率是多少?”
“Bhoonidhi 有雷达卫星吗?”
搜索场景
“查找 2024 年 1 月 Shillong 上空的 Sentinel-2 场景。”
“显示 2024 年上半年 Bengaluru 周边 20 公里范围内的 Cartosat 影像。”
“2024 年 3 月 Sundarbans 上空有 Sentinel-1 场景吗?”
“去年冬天 Kaziranga 国家公园有哪些 Landsat-8 影像覆盖?”
“查找 2023 年 12 月 Kutch 荒漠上空的 MODIS 场景。”
检查可下载内容
“在这些 Sentinel-2 场景中,我现在实际能下载多少?”
“哪些场景需要订购或付费?”
预览下载
“预览下载这些场景会获取什么。”
保存搜索以便复用
“保存这个 Sentinel-2 搜索,以便我稍后下载。”
“列出我保存的搜索。”
“显示我保存为 的搜索内容。”
“删除我保存为 的搜索。”
下载和购物车(需要登录——见下文)
“我登录 Bhoonidhi 了吗?”
“下载我保存的搜索 中的开放数据场景。”
“那个下载进行得怎么样了?”
“下载 ,完成后告诉我——我会继续做别的事。”
“把 中的付费场景加入我的购物车。”
“这周我的购物车里有什么?”
登录(用于下载和购物车)
搜索、保存的查询和预览不需要任何凭据。下载场景和将其加入购物车则需要。只需在带外登录一次——服务器会复用 bhd CLI 写入的同一个会话:
bhd auth loginMCP 服务器绝不会将用户名或密码作为工具参数接收,auth_status 也绝不会返回你的令牌。对于无法进行交互式登录的无头环境,你可以在服务器的环境中设置 BHOONIDHI_USERNAME / BHOONIDHI_PASSWORD;服务器仅读取它们以建立会话。
配置
通过环境变量设置(均为可选):
变量 | 默认值 | 用途 |
|
| 发送给 Nominatim 的 User-Agent(其使用政策要求提供描述性的 User-Agent)。 |
|
| 卫星名称匹配必须达到的置信分数(0–100);低于此分数时,返回候选列表供代理确认。 |
|
|
|
|
| 所有下载写入的允许根目录,路径为 |
|
| 并行下载工作线程数。 |
|
| 一旦下载的实时字节总数(或已知大小)超过此值,状态会将其标记为大型下载,并引导代理移交任务或运行独立命令。 |
| (未设置) | 可选的无头登录。优先使用 |
开发
uv sync
uv run pytest # test suite
uv run ruff check . # lint许可证
MIT——参见 LICENSE。
Available Tools
15 toolsauth_statusA
Report whether a Bhoonidhi login is configured for downloads and cart.
Never asks for or returns a password or token. If credentials are set in the server's environment (BHOONIDHI_USERNAME / BHOONIDHI_PASSWORD) it establishes the session so the answer matches what a download or cart action would find. Returns authenticated=True with the username when a usable session exists, or authenticated=False with guidance to log in ('bhd auth login' out of band, or set those environment variables). Call this before download or cart actions to tell the user if a login is needed.
Inputs: no named parameters.
cart_addA
Stage a saved query's scenes to the Bhoonidhi cart.
Give the slug from save_query or list_queries. Each scene is routed to the cart its access type needs (ready / on-order / priced); select narrows to specific scenes (1-based indices or scene IDs). Needs a login (see auth_status). Use this for on-order and priced scenes; priced ones still need purchasing on the portal afterwards. Returns counts of what was staged and what failed.
Inputs: slug (required), select.
cart_listA
List scenes currently staged in the Bhoonidhi cart.
Cart items are filed by the date they were added; with no window this shows today only, so pass since/until (ISO dates, e.g. "2026-08-10") or last (e.g. "1 week") to widen it. filter_by limits to a state: ready, archived, onorder, or priced. Needs a login (see auth_status).
Inputs: last, since, until, filter_by.
cart_removeA
Remove scenes from the Bhoonidhi cart.
Two ways to address rows: pass slug to index a saved query's scenes, or omit it and let select index the merged cart itself (the same row numbers cart_list shows under the same since/until/last/filter_by window). Needs a login (see auth_status).
Inputs: last, slug, since, until, +2 more.
download_queryA
Download a saved query's open-access scenes in the background.
Give the slug from save_query or list_queries. Downloads run to a fixed, server-configured root (BHOONIDHI_MCP_DOWNLOAD_ROOT, default ~/Downloads), under a per-slug folder — you cannot choose an arbitrary path. select narrows to specific scenes (1-based indices or full scene IDs); omit it for the whole query. force re-downloads files already present.
Needs a login (see auth_status). Priced and on-order scenes are skipped — stage those with cart_add instead. Returns immediately with a job_id: the download runs on its own and does NOT depend on this conversation, so never block by sleeping and re-polling. To follow it hands-free, delegate a background watcher that loops download_wait on the job_id and reports back, keeping you free to keep talking; the result's 'handoff' note says so. File sizes are unknown until each transfer starts (the portal reveals them only then); once a download proves large, download_status/download_wait flag it and 'large_download' offers a standalone command that outlives this session. Interrupted downloads restart from scratch (no resume support).
Inputs: slug (required), force, select.
download_statusA
Check a background download started by download_query (one-off).
Give the job_id from download_query. Returns the live state: running (with bytes_downloaded, mb_downloaded, rate_mb_s, percent when the total size is known, and per-scene detail), completed (with per-scene outcomes), or failed (with the error). Use this for a single progress check. To follow a job to completion without tying up the conversation, use download_wait from a delegated watcher instead. Jobs exist only while the server runs; an unknown id returns status="not_found".
Inputs: job_id (required).
download_waitA
Wait for a background download to finish, then report — for a watcher.
Give the job_id from download_query. Blocks inside the server and returns as soon as the download completes or fails, or after timeout_s (capped at 120s) with the latest progress if still running. This is the efficient way to follow a job: a delegated background watcher calls it in a loop and stops when status is "completed" or "failed", so the main conversation is never blocked on sleeps. Prefer this over repeated sleep+download_status. An unknown id returns status="not_found".
Inputs: job_id (required), timeout_s.
list_archiveA
List every satellite and sensor the Bhoonidhi portal supports.
Returns the vocabulary of valid satellites, sensors, and exact search tokens, with each product's resolution and date coverage. Call this to discover what can be searched. Set refresh=True to bypass the local cache.
Inputs: refresh.
list_queriesA
List every saved query as compact summaries.
Returns each saved query's slug, name, date range, satellites, area of interest, scene count, and a plain-English availability summary — but not the full scene lists (call show_query for one query's scenes). Use this to find the slug for a query the user saved earlier.
Inputs: no named parameters.
preview_downloadA
Dry-run a download for a search: show what would be fetched, no login.
Takes the same arguments as search_scenes, plus out_dir (where files would go) and force (preview re-downloading files already present). It runs the search and predicts, per scene, what a real download would do: would_download (staged, ready), may_404 (open data but archived — attempted but may fail until requested on the portal), already_here / already_elsewhere (a matching file exists), or skipped_on_order / skipped_priced (needs the portal).
Use this before telling a user to download, so they know how many scenes are actually fetchable. Nothing is downloaded and no login is used. File sizes are not known until a download starts (the portal exposes them only in the download response headers), and interrupted downloads cannot be resumed — both are stated in the result's disclaimers.
Inputs: end_date (required), satellite (required), start_date (required), lat, +10 more.
remove_queryA
Delete a saved query by slug.
Give the slug from save_query or list_queries. Removes the saved query from disk; the scenes themselves are unaffected. Returns status="not_found" if no query has that slug.
Inputs: slug (required).
resolve_locationA
Resolve a place name to a centroid and bounding box.
Turns a place like "Shillong" or "Loktak Lake" into latitude/longitude and a bounding box (minx, miny, maxx, maxy) that search_scenes can use as its area of interest. Returns found=False when the place can't be resolved.
Inputs: name (required).
save_queryA
Persist a search as a saved query and return a reusable slug.
Takes the same arguments as search_scenes, plus an optional name and description. Unlike search_scenes (which is stateless and leaves nothing behind), this saves the search on the portal so it can be acted on later: the returned slug is what downloading and cart staging key off. Call this once the user has confirmed a search returns the scenes they want, then hand the slug to the bhd CLI (download / cart) until those actions land in-server.
Returns status="ok" with the slug and the shaped saved query. If the satellite is ambiguous or the request is invalid, returns the same status="ambiguous_satellite" / "invalid_request" shapes as search_scenes, and saves nothing.
Inputs: end_date (required), satellite (required), start_date (required), lat, +10 more.
search_scenesA
Search Bhoonidhi scenes for a satellite over an area and date range.
The satellite may be a casual name ("Sentinel-2", "cartosat"); it is matched to the portal's exact tokens, and a constellation expands to all its platforms. Dates are ISO (YYYY-MM-DD). Give the area either as a bounding box (minx/maxx/miny/maxy) or a point with radius (lat/lon/radius_km) — typically from resolve_location. sensor narrows to one sensor on the matched satellite(s) (e.g. "SSAR", "LISS3"); product further narrows to one product under that sensor (e.g. "GCOV", "L2C-Chlorophyll") — see list_archive for the exact sensor/product names each satellite carries. The search is stateless and needs no login.
If the satellite name is ambiguous, returns status="ambiguous_satellite" with candidate names instead of guessing.
Each scene carries an "availability": Ready (downloadable now), Archived (open data but may need a portal request first), OnOrder (must be requested), or Priced (must be purchased). The result includes a plain-English "summary" of these counts and a "how_to_act" block. Tell the user clearly when scenes are Archived, OnOrder, or Priced and what each needs. This search is stateless: to act on these scenes, call save_query with the same arguments to persist them and get a , then download_query (open data) or cart_add (on-order / priced) on that slug — both need a login (see auth_status). Downloads cannot be resumed if interrupted (the portal has no range support).
Inputs: end_date (required), satellite (required), start_date (required), lat, +8 more.
show_queryA
Return one saved query by slug, with its scenes.
Give the slug from save_query or list_queries. Returns the full saved query: its selections, area of interest, date range, and shaped scenes with availability. Returns status="not_found" if no query has that slug.
Inputs: slug (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 15 tools
Each tool targets a distinct action or resource: search is separated from preview, save, download, and cart operations, and download_status vs download_wait are clearly one-off vs watcher. Even where arguments overlap, the descriptions draw explicit boundaries.
Most query tools follow verb_noun (remove_query, search_scenes, list_queries), but cart_* uses noun-verb order (cart_add, cart_list) and auth_status/download_status/download_wait are noun-phrase style. Everything is snake_case and readable, but the naming pattern is mixed rather than uniform.
15 tools is within the ideal range, and each tool covers a distinct stage of the archive-to-download/cart workflow. There is no apparent redundancy or padding.
The set covers discovery, geocoding, search, dry-run preview, saved-query lifecycle, auth status, download execution/monitoring, and cart management. Minor gaps like no in-server edit of saved queries, no active-download listing, and no direct archival request action are workarounds via the portal or CLI.
Maintenance
Related MCP Connectors
Geospatial AI MCP server — satellite imagery, embeddings, weather, GNS governance
MCP server connecting AI agents to 100+ apps (Gmail, Slack, Notion, GitHub) via one-click OAuth.
One MCP server for the services your AI needs: memory, files, Microsoft 365, kanban and more.
MCP server for building and testing AI agents with multi-model experimentation and insights.
Related MCP Servers
FlicenseBqualityDmaintenanceAn MCP server that integrates AI retrievals with NASA's Common Metadata Repository (CMR), allowing users to search NASA's catalog of Earth science datasets through natural language queries.14-- AlicenseNot gradedqualityDmaintenanceMCP server for interacting with Google Earth Engine, enabling geospatial analysis such as dataset visualization, statistics computation, and search via AI assistants.16MIT
- AlicenseAqualityDmaintenanceA Python MCP server that provides unified access to satellite and geospatial data through natural language queries, with automatic place name geocoding and support for raster, vector, and Zarr formats.24Apache 2.0
- AlicenseNot gradedqualityAmaintenanceAn MCP server that connects AI agents to cloud-native geospatial data via STAC metadata and DuckDB with H3 spatial indexing, enabling zero-configuration SQL queries on terabyte-scale datasets over S3.22BSD 3-Clause