expense-mcp
A minimal FastMCP server providing demo tools for arithmetic operations and dice rolls, designed for learning MCP wiring with Claude Desktop or other MCP clients.
README
expense-mcp
A minimal FastMCP server that exposes demo tools for arithmetic and dice rolls. Use it to learn MCP wiring with Claude Desktop or other MCP clients.
Requirements
- Python 3.13+ (see
.python-version/pyproject.toml) - uv for environments and dependency management
Setup with uv (step by step)
Follow these in order on a machine that does not have the project set up yet.
1. Install uv
Pick the official installer for your OS. Common one-liners:
Windows (PowerShell):
irm https://astral.sh/uv/install.ps1 | iex
macOS / Linux:
curl -LsSf https://astral.sh/uv/install.sh | sh
Close and reopen the terminal (or reload your shell profile) so uv is on your PATH.
2. Confirm uv
uv --version
3. Get the code
Clone the repository and move into it (use your real clone URL or path):
git clone <repository-url> expense-mcp
cd expense-mcp
If you already have the folder unpacked, just cd to the repo root (where pyproject.toml lives).
4. (Optional) Install a matching Python with uv
If you do not have Python 3.13 available, uv can fetch one:
uv python install 3.13
The repo’s .python-version pins 3.13; uv sync will respect that when creating the environment.
5. Create the virtual environment and install dependencies
From the repo root:
uv sync
This will:
- Create
.venvif it is missing (or update it to match the lockfile) - Install packages from
pyproject.tomlusing versions fromuv.lock
If you prefer to create the venv yourself first (optional):
uv venv --python 3.13
uv sync
6. (Optional) Activate the virtual environment
You do not need activation if you always use uv run … (recommended). To activate:
- Windows (PowerShell):
.venv\Scripts\Activate.ps1 - Windows (cmd):
.venv\Scripts\activate.bat - macOS / Linux:
source .venv/bin/activate
7. Verify the install
uv run fastmcp --version
uv run python --version
You should see a FastMCP version and Python 3.13.x.
Quick reference (after setup)
| Goal | Command |
|---|---|
| Reinstall deps from lockfile | uv sync |
| Add a new dependency | uv add <package> |
Refresh uv.lock after editing pyproject.toml |
uv lock |
| Run a one-off command in the project env | uv run <command> |
Run the server
Development / inspector (useful for debugging tools in the browser):
uv run fastmcp dev inspector main.py
Production-style run (stdio transport for MCP clients):
uv run fastmcp run main.py
Direct Python (stdio, same entry as if __name__ == "__main__"):
uv run python main.py
Tools (API)
Defined in main.py:
| Tool | Description |
|---|---|
roll_dice |
Rolls n_dice six-sided dice; returns integers. |
add_numbers |
a + b |
subtract_numbers |
a - b |
multiply_numbers |
a * b |
Docstrings on each function are what MCP clients show as tool descriptions.
Install into Claude Desktop
uv run fastmcp install claude-desktop main.py
Then:
- Restart Claude Desktop fully (quit the app, not only the window).
- In settings, ensure the expense-mcp (or similarly named) server is enabled.
- If the install command errors (especially on Windows), see Claude Desktop: config path / install fails.
Troubleshooting
Use this section when something fails or behaves unexpectedly. Work from the top of each subsection that matches your symptom.
uv and Python
| Symptom | What to try |
|---|---|
uv is not recognized |
Install uv (see Setup), then restart the terminal. Confirm install docs for your OS put uv on PATH. |
uv sync fails or wrong Python |
Run uv python install 3.13, then uv sync again from the repo root (folder that contains pyproject.toml). |
| Packages seem missing | From repo root: uv sync. Prefer uv run <command> so the project .venv is used; avoid mixing in a global python without the venv. |
“No pyproject.toml” or wrong deps |
cd to the repository root before uv sync / uv run. |
Windows: PowerShell and scripts
| Symptom | What to try |
|---|---|
Running .venv\Scripts\Activate.ps1 fails (“running scripts is disabled”) |
For your user only: Set-ExecutionPolicy -Scope CurrentUser RemoteSigned in PowerShell, then retry. Or use cmd and .\venv\Scripts\activate.bat. |
| Activation not needed | Using uv run … from the repo root avoids activation entirely. |
Server / dev inspector
| Symptom | What to try |
|---|---|
ModuleNotFoundError or wrong fastmcp |
Run with uv run (e.g. uv run fastmcp dev inspector main.py), not a system-wide python that never had deps installed. |
| Inspector / dev command errors | Run from repo root; pass main.py or a path that exists. If a port is busy, close other inspector sessions or see FastMCP docs for port options. |
| Stdio server exits immediately | In a terminal: uv run fastmcp run main.py or uv run python main.py and read the traceback. Fix any import or config error before wiring Claude again. |
Claude Desktop: MCP does not connect or server errors
| Symptom | What to try |
|---|---|
| Server disabled | Settings → MCP: turn on the expense-mcp entry; restart Claude. |
| Wrong working directory | The MCP process must run with cwd (or equivalent) set to this repo so main.py and .venv resolve. Re-run uv run fastmcp install claude-desktop main.py or edit claude_desktop_config.json so the launched command’s working directory is the project root. |
uv not found when Claude starts |
Use the full path to uv.exe in config, or ensure the user PATH seen by the desktop app includes uv (often differs from your interactive terminal). |
| JSON config mistakes | Open claude_desktop_config.json in an editor with JSON validation. One missing comma breaks the whole file. |
Manual config shape (illustrative) — names and keys follow your Claude Desktop version; adjust paths:
"mcpServers": {
"expense-mcp": {
"command": "C:\\Users\\you\\AppData\\Local\\Programs\\uv\\uv.exe",
"args": ["run", "fastmcp", "run", "main.py"],
"cwd": "C:\\path\\to\\expense-mcp"
}
}
On macOS/Linux use forward slashes and typical uv location from which uv. Re-installing via fastmcp install is usually easier than hand-editing.
Claude Desktop: config path or install fails (Windows Store and paths)
Classic path (many installers): %APPDATA%\Claude (contains claude_desktop_config.json).
Microsoft Store build often uses:
%LOCALAPPDATA%\Packages\Claude_<id>\LocalCache\Roaming\Claude
If fastmcp install claude-desktop says Claude has no config folder or cannot find it:
-
Environment variable — set the directory that actually holds
claude_desktop_config.json(names depend on FastMCP version):FASTMCP_CLAUDE_DESKTOP_CONFIG_DIR, orCLAUDE_DESKTOP_CONFIG_DIR
Then runuv run fastmcp install claude-desktop main.pyagain in the same terminal session.
-
Patch in
.venv(advanced) — Patchingfastmcp\cli\install\claude_desktop.pyunder.venvcan teach the installer to scanPackages\Claude_*\LocalCache\Roaming\Claude. Any lateruv syncor fresh venv removes that patch — re-apply or use option 1 or 3. -
Directory junction — If
%APPDATA%\Claudedoes not already exist as a real folder, junction it to the Store path (adjust<your-id>; may need an elevated cmd):mklink /J "%APPDATA%\Claude" "%LOCALAPPDATA%\Packages\Claude_<your-id>\LocalCache\Roaming\Claude"Then run
uv run fastmcp install claude-desktop main.pyagain.
Upstream FastMCP may add first-class support for the Store layout so patching or junctions are unnecessary.
Assistant answers without using your tools
A normal-looking answer (e.g. “8 + 10 = 18”) does not prove the MCP ran. Check for tool-use in the UI or force a unique tool.
- UI — Look for a tool / “used …” step, not only the final message.
- Explicit instruction — e.g. “Use the
add_numberstool from expense-mcp for 8 and 10 and state that you used the tool.” - Unique tool — e.g. “Call
roll_dicewithn_dice=4.” Matching tool name and die results implicates this server (unless another MCP defines the same tool name). - Proof tool — Add a tiny tool in
main.pythat returns a string only this project would emit; call it once and compare the tool output.
Patches or local edits under .venv disappeared
Anything changed only inside .venv (including hand-patched FastMCP files) is replaced when you run uv sync, delete .venv, or reinstall deps. Prefer env vars, junctions, or documented config over long-lived venv edits; otherwise re-apply the patch after each sync.
Project layout
| Path | Role |
|---|---|
main.py |
FastMCP server and @mcp.tool definitions |
pyproject.toml |
Package metadata and dependencies |
uv.lock |
Locked dependency versions |
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。