MCP Smoke Test

MCP Smoke Test for Stata-MCP

This skill performs an end-to-end smoke test of the local Stata-MCP server. It checks installation, runs representative data and Stata commands, and reports pass/fail status for each tool.

When to use

• The user asks whether Stata-MCP is working.
• The user wants to verify a new MCP server configuration.
• The user says a tool is broken or unavailable.
• The user asks for a smoke test, sanity check, or end-to-end test.

Before you start

1. Make sure the current working directory is the project root.
2. Confirm whether the project virtual environment exists at ./.venv/bin/stata-mcp. If it does not exist, use the globally installed stata-mcp command instead.
3. Ensure ~/.statamcp/debug.toml exists and enables beta security. Use .claude/skills/mcp-smoke-test/config/debug.example.toml as a template.

Step 1: Ensure the MCP server is registered

Check whether the MCP server stata-mcp-local-smoke-test is available by listing Claude MCP servers:

claude mcp list

If stata-mcp-local-smoke-test is missing, add it with the debug configuration:

claude mcp add stata-mcp-local-smoke-test -s local -- $(pwd)/.venv/bin/stata-mcp -c ~/.statamcp/debug.toml

If ./.venv does not exist, use:

claude mcp add stata-mcp-local-smoke-test -s local -- stata-mcp -c ~/.statamcp/debug.toml

After adding, if the server is still not available, tell the user:

MCP server was added but is not yet visible. Please restart the Claude process and then invoke the smoke test again.

Step 2: Run smoke tests

Run the preparation script. It clears cached summaries, locates or generates a suitable auto.dta, and copies the boundary test files to /tmp:

bash .claude/skills/mcp-smoke-test/scripts/prepare_smoke_test.sh

Then use the mcp__stata-mcp-local-smoke-test tools (or the server name registered above) to execute each test below. Record the result for each item.

For concrete tool-call parameters, expected outputs, and troubleshooting, see .claude/skills/mcp-smoke-test/examples.md.

Test A — get_data_info on a local Stata sample file

Try the first existing file from this list:

• /Applications/Stata/auto.dta
• /Applications/StataNow/auto.dta

Expected result: JSON summary with 74 observations and 12 variables.

Test B — get_data_info on an allowed URL

Use this URL (already in the allowlist):

https://raw.githubusercontent.com/mwaskom/seaborn-data/master/iris.csv

Expected result: JSON summary with 150 observations and 5 variables.

Test C — stata_do with a legal dofile

Run the bundled dofile .claude/skills/mcp-smoke-test/scripts/legal.do with stata_do.

Expected result: execution succeeds and returns log file paths.

Test D — stata_do security boundary (negative test)

Run the copied dofile /tmp/mcp_smoke_test_boundary.do with stata_do.

Expected result: blocked by the security guard because the dofile is outside the allowed working directory.

Test E — read_log on the generated log

Read the text log returned by Test C. Expected result: log content includes regression output.

Test F — help for a Stata command

Run help with argument regress. Expected result: Stata help text for regress.

Test G — ado_package_install (optional, requires user approval)

If the user confirms, attempt to install a small SSC package such as estout. Expected result: user approval prompt, then installation log.

If the user declines or the tool is not registered (non-unsafe profile), mark this test as skipped.

Step 3: Clean up

Run the cleanup script to remove temporary files and cached summaries:

bash .claude/skills/mcp-smoke-test/scripts/cleanup.sh

Do not delete log files or data summaries; they are useful for diagnosis.

Step 4: Report results

Return a concise Chinese report using this template:

# Stata-MCP 冒烟测试报告

## 服务器状态
- 注册状态：<已注册 / 新注册 / 需重启>
- 配置文件：`~/.statamcp/debug.toml`

## 测试结果
- [ ] Test A 本地 auto.dta 读取：<通过 / 失败 / 跳过> — 原因
- [ ] Test B URL 鸢尾花数据集读取：<通过 / 失败 / 跳过> — 原因
- [ ] Test C 合法 dofile 执行：<通过 / 失败 / 跳过> — 原因
- [ ] Test D 安全边界拦截：<通过 / 失败 / 跳过> — 原因
- [ ] Test E 日志读取：<通过 / 失败 / 跳过> — 原因
- [ ] Test F help 命令：<通过 / 失败 / 跳过> — 原因
- [ ] Test G ado 包安装：<通过 / 失败 / 跳过> — 原因

## 总结
<一句话结论：全部通过 / 部分失败 / 需要用户操作>

Only mark a test as passed when you observed the expected result. For failures, include the exact error message or tool response.

Resources