@getfacade/mcp
Enables an AI agent to create a building from a photo, generate a facade design, and order a line-by-line cost estimate and documentation album, using country-specific materials and construction norms.
README
@getfacade/mcp
MCP server for GetFacade.ai: an agent creates a building, uploads a photo of it, gets the exterior designed, priced and documented, without a human in front of a screen.
Each design is worked out for the country the building stands in: materials that are applicable there, manufacturer products that are actually sold there, and the technical build-up behind the surface. The render shows that on the photo of the house.
The estimate and the album come from the same design:
- The estimate prices it line by line, in materials and labour, at what those materials cost in that country.
- The album documents it for the crew that builds it: the build-up of the facade, safety notes and the norms behind them.
The server is a thin wrapper over the public GetFacade API. It stores nothing, caches nothing and decides nothing: every rule (pricing, admission, the colour grammar, every message) stays on the server, and the wrapper only carries calls and answers.
Requirements
- Node.js 20+
- An agent API key. Issue one for yourself at app.getfacade.ai → Account → Settings → API. The value is shown once and cannot be recovered; each key carries a hard spend cap.
- Agent access is paid. There are no trial credits in it: the key spends from the agent wallet, and without a balance the first paid call is refused with an explanation. An active Pro Plan fills that wallet up to 1,000 credits once per billing period; past that, credits are bought.
Setup
Claude Desktop / any MCP client, mcpServers section:
{
"mcpServers": {
"getfacade": {
"command": "npx",
"args": ["-y", "@getfacade/mcp"],
"env": { "GETFACADE_API_KEY": "your-key" }
}
}
}
To run a checkout instead of the published package, point the client at it:
"command": "node", "args": ["/path/to/mcp/src/index.js"].
| Variable | Required | Default |
|---|---|---|
GETFACADE_API_KEY |
yes | — |
GETFACADE_API_BASE_URL |
no | https://api.getfacade.ai/api/v1 |
GETFACADE_LANG |
no | en |
Messages, including refusals, come from the service in the language of
GETFACADE_LANG. Set it to the language the answer should be read in.
Tools
| Tool | What it does |
|---|---|
create_building |
Creates a building. The name is unique per account. |
upload_photo |
Registers a view, uploads the bytes, confirms, waits for validation. |
start_design |
Creates a design and queues its render. Returns a job id. |
refine_design |
Changes a finished design in words. Every step after the first. |
get_job |
Polls one render, estimate or album. |
list_jobs |
Recent jobs across the account, unfinished first. |
list_designs |
Designs of a building with their renders and finished picture. |
order_estimate |
Orders a cost estimate for chosen renders. |
order_album |
Orders the PDF album for chosen renders. |
upscale_render |
Enlarges a finished render. Costs tokens, asynchronous. |
get_estimate |
The estimate itself: totals, assumptions and every line. |
add_estimate_line |
Adds one line to an estimate. |
update_estimate_line |
Edits one line of an estimate. |
delete_estimate_line |
Removes one line from an estimate. |
delete_render |
Deletes one render. The main render's design returns to draft. |
delete_design |
Deletes one design with the renders under it. |
delete_building |
Deletes a building with everything under it. |
list_token_packages |
The packages this account can buy. |
buy_tokens |
Refills this key's wallet. See below. |
get_balance |
Agent wallet, key cap, and whether the next paid call will be accepted. |
report_problem |
Reports a defect in this API. Free, and works on an empty wallet. |
Rendering is asynchronous: start_design, refine_design, order_estimate and order_album return a job id
immediately, and get_job reports when it is done. A finished estimate reads ready where a
render and an album read completed. The only call that waits is upload_photo, which polls
until the photo is accepted or rejected.
A paid call is never ordered twice
start_design, refine_design, order_estimate, order_album and upscale_render cost
money, and the charge follows the job that gets created. If one of them times out, the honest
question is whether the job exists, and the agent cannot answer it from where it stands.
So this server answers it instead. Every paid call goes out under a name of its own
(Idempotency-Key), and this server, not the agent, retries it under that same name when the
connection fails or when the API says the first attempt is still running. The API recognises
the repeat and hands back the original job, so a timeout costs one design, not two.
Nothing about this reaches the tool arguments: ordering the same design twice on purpose stays possible, because two calls are two names. What is gone is the accidental second order.
The API requires that name on every paid call made with an API key, so calling it without this wrapper means sending the header yourself. The reference is at getfacade.ai/agents.
What the key can reach
A key is a bearer secret that lives in your agent's config, so its reach is the published agent
surface and nothing else: buildings, photos, designs, renders, estimates, albums, the jobs list
and the API wallet. Anything outside that answers 403 AGENT_SURFACE_FORBIDDEN, including the
account's own sign-in settings, its subscription and its payment history. Changing those, and
issuing or revoking keys, is done by a person signed in to the app.
The reference is at getfacade.ai/agents.
Results are permanent public links
A finished render (result_url, main_render_url) and a finished album (result_url) are
served from stable URLs: no signature, no expiry. An assistant can hand one straight to a
person as the answer to "show me the result", with nothing to refresh and no second call.
The honest half of that: because the link is unsigned it asks nobody for permission, so it
keeps working for whoever it is forwarded to, and it cannot be recalled afterwards. Share it as
deliberately as any other link that is public forever. GET /renders/{render}/download is a
different thing — a short-lived signed URL with a filename, for saving the file rather than for
sharing it.
Deletions are soft on the server: a person can undo them in the app, this server cannot.
buy_tokens works only if the key was issued with purchasing enabled (a switch on the
issue screen), and only up to what the key may still spend — a purchase never lifts the key's
own spend cap. It answers charged when the payment provider took it from a saved payment
method, or requires_human with a checkout_url for a person to open. Either way the tokens
are credited when the payment is confirmed, so poll get_balance.
Issuing keys and buying a subscription are deliberately absent: a key that can issue itself a key has no ceiling, and a subscription is bought by a person, in the app.
License
MIT (see LICENSE) — this wrapper carries calls and answers and nothing else, so there is
nothing in it to keep closed. The GetFacade service it talks to is a separate matter: using it
is governed by the terms at getfacade.ai/terms, and a key is
issued to an account, not to this package.
Development
npm install
npm test # smoke test against a stub of the API
test/live-journey.mjs drives the whole journey through a real MCP client against a real API
with a real key. It is not part of npm test, because it spends actual credits; the header of
the file says how to run it.
Releases: see RELEASING.md.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。