接入你的智能体
从发布者处安装。不是 PlugMCP 端点。
远程 URL 还是本机命令? 两种安装方式的区别
Quick Start · Features · Security · Examples · Contributing
Stop hand-writing MCP servers. Point
mcpfoundryat a database or an OpenAPI spec and get a clean, runnable, secure-by-option MCP server whose tools actually run — real parameterised SQL or real HTTP calls, notTODOstubs. Node.js or Python. Zero boilerplate.
npx mcpfoundry create \
--type openapi \
--input https://petstore3.swagger.io/api/v3/openapi.json \
--output ./petstore-mcp
cd petstore-mcp && npm install && npm start
# 🎉 MCP server live on http://localhost:3000/mcp
That's a full MCP server — every endpoint turned into a validated tool — running. No SDK wrangling, no transport plumbing, no boilerplate.
Connect it to Claude in one step: every generated project includes a ready-to-use .mcp.json. Open the folder in Claude Code (it auto-detects the file), or paste the block into Claude Desktop's config. The generated README.md has the exact snippet.
Every endpoint becomes a typed, validated MCP tool.
mcpfoundry create --type openapi --input ./openapi.yaml --output ./my-server
Tables are introspected into CRUD tools. SQLite tools run real, parameterised SQL out of the box — just point DATABASE_PATH at your file:
mcpfoundry create --type database --provider sqlite --uri ./app.db --output ./db-server
cd db-server && npm install && DATABASE_PATH=../app.db npm start
Postgres is also introspected (its handlers are scaffolded stubs for now):
mcpfoundry create --type database \
--provider postgres \
--uri "postgresql://user:pass@localhost:5432/mydb" \
--output ./pg-server --lang python
💡 Preview first with
--dry-runto see exactly which tools you'll get — no files written:✔ Dry run — 4 tool(s) would be generated: • list_pets(limit?: integer) • create_pet(name: string, tag?: string) • get_pet_by_id(pet_id: integer) • delete_pet(pet_id: integer)
Pass --secure and every generated server enforces zero-trust access — recommended, never required:
Authorization: Bearer header per request over HTTP, or ZTAI_AUTH_TOKEN at startup over stdio) against JWT_SECRET. Invalid or missing → rejected before any tool runs.--secure — it's just good hygiene).ZTAI_CANARY_ID and tool output carries a subtle, traceable marker to help detect adversarial data exfiltration.Without --secure you still get a perfectly good, vendor-neutral MCP server.
SQLite & Postgres are introspected today; SQLite emits working SQL. MySQL & MongoDB are stubbed and open for contributions.
mcpfoundry follows a clean Template-Compiler pattern:
source (DB / OpenAPI) ──▶ parser ──▶ normalized IR (ToolSpec[])
│
Handlebars compiler
│
▼
templates/<lang>/ ──▶ your generated server
Parsers and templates are decoupled by a normalized intermediate representation, so adding a new language is just a new templates/<lang>/ folder — no engine changes. See CONTRIBUTING.md.
Two of the most common contributions — a new language template or a new database provider — need zero changes to the core engine. See CONTRIBUTING.md.
MIT — build freely.
⭐ Star on GitHub if mcpfoundry saved you an afternoon.
Forge production-ready MCP servers from databases or OpenAPI specs.
工具
暂无工具列表。远程服务器在首次健康检查后会获得工具列表。
常见问题
在本页复制适用于 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 配置。它使用发布者的安装方式,而不是 PlugMCP 代理。
此记录未列出必需的环境变量。
Mcpfoundry 通过 stdio 在本地运行。
上次检查时间:2026-10-28。 Forge production-ready MCP servers from databases or OpenAPI specs.