swag
Enables agents to browse a catalog of OpenAPI specs, search for operations, and retrieve full operation contracts to build API requests without calling the target APIs.
README
swag
MCP server (Python, FastMCP).
Run
The only runtime mode is an HTTP server: FastAPI + MCP on /mcp and GET /health. Locally it runs via Docker; in production the same image runs (uv run uvicorn swag.app.asgi:app).
Local (Docker)
docker compose up --build
Compose mounts ./swag into the container and runs uvicorn with --reload, so code changes are picked up without rebuilding the image.
Check:
curl http://localhost:8000/health
# {"status":"ok","service":"swag"}
# MCP for Cursor: http://localhost:8000/mcp
Cursor
Connect to the running server over HTTP:
{
"mcpServers": {
"swag": {
"url": "http://localhost:8000/mcp"
}
}
}
Tools
| Tool | Description |
|---|---|
list_services |
Returns id, name, description for each registered API (no OpenAPI bodies or spec URLs) |
search_spec |
Searches a selected service spec and returns compact ranked operation hits (method, path, summary, score) |
get_operation |
Returns one operation's full contract for a chosen method+path: parameters, request body, and responses, with local $ref schemas resolved inline |
On MCP initialize, the server sends instructions telling the agent to call list_services first, pick a service_id from name/description, then search_spec with the user's action/entity query and any clear method/path/tag hints, and finally get_operation on the chosen hit to fetch the details needed to build a request.
search_spec does not return the full OpenAPI document. It builds an in-memory search index from the selected spec (JSON or YAML) and returns top operation candidates; get_operation then returns the self-contained contract of a single operation. The server never calls the target API itself.
Configuration
| Variable | Default | Description |
|---|---|---|
SWAG_HOST |
0.0.0.0 |
HTTP bind host |
SWAG_PORT |
8000 |
HTTP port |
SWAG_MCP_MOUNT_PATH |
/mcp |
MCP Streamable HTTP mount path |
SWAG_CATALOG_PATH |
data/catalog.json |
Path to the services catalog JSON |
SWAG_SPEC_FETCH_TIMEOUT |
15.0 |
HTTP timeout (seconds) for fetching a spec document |
Layout
The code is organized into vertical feature packages: each package is
self-contained (its own models / service / tool).
swag/
config.py # settings
exceptions.py # SwagError hierarchy
mcp_instructions.py # server instructions for agents
catalog/ # services registry (list_services): models, source, service, tool
spec/ # raw OpenAPI/Swagger document: models, fetch, decode, validate, parsing, service
search/ # search index + ranking (search_spec): models, text, extractors,
# index, keyword, fuzzy, boosters, engine, tool
operation/ # full contract of a single operation (get_operation): models, detail, ref_resolver, tool
app/ # composition root: gateway (orchestrator), tools, server (FastMCP), asgi (FastAPI)
data/
catalog.json # production catalog (URLs only)
Development
Install dev dependencies and pre-commit hooks (runs ruff + mypy from pyproject.toml before each commit):
uv sync --group dev
task pre-commit-install
# or: uv run pre-commit install
| Task | Description |
|---|---|
task test |
Run pytest |
task up |
docker compose up --build |
task down |
docker compose down |
task ruff |
Ruff check (no auto-fix) |
task ruff-fix |
Ruff check with --fix |
task mypy |
Mypy type check |
task check |
Ruff + mypy |
task pre-commit-install |
Install git hooks |
Run all hooks manually:
uv run pre-commit run --all-files
Requires go-task: brew install go-task (or see project docs).
Tests
uv run pytest
# or: task test
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。