mcp.andreglegg.no
Remote MCP tool host serving tree generation tools over Streamable HTTP, returning short-lived download URLs for artifacts.
README
mcp.andreglegg.no
Remote MCP tool host. Serves my own tools over Streamable HTTP so any MCP client can use them without a local install.
| Route | Auth | Tools |
|---|---|---|
/treegen |
none | generate_tree, random_tree, list_presets |
/assetcut |
Cloudflare Access service token | ping only — real tools pending |
/dl/<id> |
none | artifact downloads, 30 min TTL |
Use it
claude mcp add --transport http treegen https://mcp.andreglegg.no/treegen
Tools return a short-lived download URL rather than the file itself — a detail-2 tree GLB is ~664 KB, which has no business inside a model's context window.
Run locally
npm install
npm test
PUBLIC_BASE_URL=http://localhost:8787 npm start
Environment: PORT (default 8787), ARTIFACT_DIR, PUBLIC_BASE_URL.
Architecture
Design and rationale — including why Cloudflare Workers was rejected and why
binaries are never returned inline — live in
docs/superpowers/specs/2026-08-03-mcp-subdomain-design.md.
The host runs on a self-hosted Windows machine behind a free Cloudflare Tunnel, so the public endpoint is only up when that machine is.
Adding a tool
- Write
src/tools/<name>/index.jsexportingregister<Name>Tools(server, { store }). - Add it to the route's array in
src/registry.js. That's the only file that changes. - Return
store.put(bytes, ext)'s URL — never raw bytes.
Deploying to a host machine
The scripts take paths as parameters and default to the current user's profile, so nothing here is tied to a particular machine.
First-time setup, from the checkout directory:
npm ci --omit=dev
powershell -File install.ps1
install.ps1 registers mcp-host as a Scheduled Task at boot. The Node path is
absolute inside it because the task runs as SYSTEM, whose PATH excludes the
Node install directory; a bare node there fails with 0x80070002 while the
task still reports Ready and nothing listens.
The tunnel. cloudflared tunnel login needs a browser and cannot be
scripted. It can run on any machine, with the resulting credentials JSON copied
to the host:
cloudflared tunnel login # writes cert.pem
cloudflared tunnel create <tunnel-name> # writes <UUID>.json
cloudflared tunnel route dns <tunnel-name> mcp.andreglegg.no
Then on the host machine, copy <UUID>.json into the cloudflared directory,
copy tunnel/config.example.yml to config.yml there and fill in the tunnel
UUID, and run:
powershell -File install-tunnel.ps1
Do not use cloudflared service install. On Windows it registers the
service with no arguments and drops the --config flag, so the service runs as
LocalSystem, finds no config, and sits there connecting to nothing while
reporting RUNNING. Copying the config into the LocalSystem profile does not fix
it either. The scheduled task in install-tunnel.ps1 pins the full command line
and works.
Redeploying after a push — run update.ps1 on the host machine:
powershell -File update.ps1
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。