mcp-forge
A FastAPI-style framework for building MCP servers in Python, enabling declarative tool definitions with automatic schema validation and multiple transport options.
README
⚡ mcp-forge
The FastAPI-style framework for building MCP servers in Python.
Stop writing boilerplate MCP servers. mcp-forge lets you declare tools with a single decorator — schema, validation, and transport are handled automatically.
from mcp_forge import Forge
app = Forge(name="my-server", version="1.0.0")
@app.tool(description="Add two numbers")
def add(a: int, b: int) -> int:
return a + b
if __name__ == "__main__":
app.run() # STDIO, HTTP or SSE — your choice
That's it. No JSON schema by hand. No transport boilerplate. No config files.
✨ Features
| Feature | Description |
|---|---|
| Declarative tools | @app.tool() — auto-infers schema from Python type hints |
| Auto schema | Pydantic v2 under the hood — full JSON Schema generation |
| Type validation | Input/output validated at runtime, errors surfaced cleanly |
| Multi-transport | STDIO (Claude Desktop), HTTP (REST), SSE (streaming) |
| Async-first | Native async def support for all tools |
| Hot reload | --reload flag for development |
| CLI | mcp-forge new, mcp-forge run, mcp-forge build |
| Zero config | Sensible defaults, override when needed |
| PEP 561 typed | Ships py.typed — full mypy/pyright support out of the box |
🚀 Quick Start
Install
pip install mcp-forge
Create a new server
mcp-forge new my-server
cd my-server
mcp-forge run --reload
Add your first tool
from mcp_forge import Forge
from pydantic import BaseModel
app = Forge(name="calculator", version="1.0.0")
class SearchInput(BaseModel):
query: str
limit: int = 10
@app.tool(description="Search the knowledge base")
async def search(params: SearchInput) -> list[dict]:
# your logic here
return [{"result": f"Result for {params.query}"}]
Run in different transports
# STDIO (for Claude Desktop, Cursor, VS Code)
mcp-forge run --transport stdio
# HTTP (for custom integrations)
mcp-forge run --transport http --port 8080
# SSE (for streaming)
mcp-forge run --transport sse --port 8080
📐 Architecture
mcp-forge
├── core/
│ ├── forge.py # Main Forge class — app entrypoint
│ ├── schema.py # Auto JSON Schema from type hints (Pydantic v2)
│ ├── validator.py # Input/output validation engine
│ ├── config.py # ForgeConfig dataclass
│ └── exceptions.py # Exception hierarchy
├── transports/
│ ├── stdio.py # STDIO transport (Claude Desktop compatible)
│ ├── http.py # FastAPI-based HTTP transport
│ └── sse.py # Server-Sent Events transport
├── cli/
│ └── main.py # Typer CLI — new, run, build, list
└── contrib/
├── memory.py # Built-in memory tool (isolated per-instance store)
├── filesystem.py # Built-in filesystem tools (read, write, list, delete)
└── web.py # Built-in web fetch tool
🔌 Built-in Tools (contrib)
mcp-forge ships with production-ready contrib tools you can include in one line:
from mcp_forge import Forge
from mcp_forge.contrib import memory, filesystem, web
app = Forge(name="full-server")
app.include(memory) # remember(), recall(), forget(), list_memory()
app.include(filesystem) # read_file(), write_file(), list_dir(), path_exists(), delete_file()
app.include(web) # fetch_url()
⚙️ Configuration
from mcp_forge import Forge, ForgeConfig
app = Forge(
name="my-server",
version="1.0.0",
config=ForgeConfig(
transport="http",
port=8080,
log_level="info",
cors_origins=["*"],
max_tool_timeout=30,
)
)
Or via mcp-forge.toml:
[server]
name = "my-server"
version = "1.0.0"
transport = "http"
port = 8080
log_level = "info"
[tools]
max_timeout = 30
auto_reload = true
🧪 Testing
from mcp_forge.testing import ForgeTestClient
client = ForgeTestClient(app)
def test_add():
result = client.call("add", {"a": 2, "b": 3})
assert result == 5
# Inside async tests:
async def test_add_async():
result = await client.acall("add", {"a": 2, "b": 3})
assert result == 5
🌍 Ecosystem Compatibility
| Client | Transport | Status |
|---|---|---|
| Claude Desktop | STDIO | ✅ Supported |
| Cursor | STDIO | ✅ Supported |
| VS Code (Copilot) | STDIO / HTTP | ✅ Supported |
| Continue.dev | HTTP / SSE | ✅ Supported |
| Custom LLM apps | HTTP / SSE | ✅ Supported |
📦 Roadmap
- [x]
@tooldecorator with auto-schema - [x] STDIO transport (MCP spec
2024-11-05) - [x] HTTP transport (FastAPI)
- [x] SSE transport
- [x] Pydantic v2 validation
- [x] CLI (
new,run,build) - [x] Contrib routers (memory, filesystem, web)
- [x] PEP 561
py.typedmarker - [x] Python 3.11 / 3.12 / 3.13 tested
- [ ]
@resourcedecorator (MCP Resources spec) - [ ]
@promptdecorator (MCP Prompts spec) - [ ] Plugin registry (community contrib tools)
- [ ] Docker one-liner deploy
- [ ] mcp-forge cloud (hosted MCP servers)
🤝 Contributing
PRs welcome. See CONTRIBUTING.md.
git clone https://github.com/nsfwbunny/mcp-forge
cd mcp-forge
pip install -e ".[dev]"
pytest
📄 License
MIT — see LICENSE.
<p align="center"> Built with ⚡ by <a href="https://github.com/nsfwbunny">Benni Alencar</a> </p>
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。