SignOff Studio

io.github.samuccaya由 GitHub 账号 samuccaya 发布v26.8.42更新于 Nov 19, 2026

Local API and database verification for AI coding agents: HTTP checks and read-only SQL.

已验证STDIO仅桌面数据库开发者工具

概览

AI 生成的概览

本地 MCP 服务器,让 AI 编程助手发送 HTTP 请求并执行只读 SQL,以证明 API 可用且数据库中的数据正确。

可以做什么
  • 发送临时 HTTP 请求,并对状态码、延迟、响应头、正文文本和 JSON 路径做内联校验
  • 按名称、URL 或文件夹搜索、读取并运行已保存的请求或整个集合,可分页查看大结果
  • 查看数据库连接与表结构,执行只读查询,并对行数、标量或列内容做断言
  • 把请求与数据库校验绑定,用一次判定验证 API 产生的效果
  • 将验证过的请求与检查保存到工作区,导入 OpenAPI、Postman、Insomnia、Bruno 或 cURL,启动本地 mock API,并导出 HTML/Markdown 报告、JUnit 与 PR 摘要
适用场景
适合需要确认接口行为符合预期、且产生的数据确实写入数据库的场景,也适合希望把检查保存下来供后续会话或同事重跑的情况。
运行要求
以 stdio 在本地运行;需要 Node.js 22 或更高版本,或使用自带同一服务器的 SignOff Studio 桌面应用。数据库密码来自连接配置的密码环境变量键,或来自正在运行的桌面应用。可选 SIGNOFF_WORKSPACE 指定工作区目录;SIGNOFF_TELEMETRY=0 或 DO_NOT_TRACK=1 可关闭遥测。无需账号或云端中转。
安装前请注意
对于所有者标记为允许代理写入的连接,db_execute 可以执行写操作,但标记为生产环境的连接会拒绝代理写入,且默认使用回滚式试运行;请先核对该设置。已保存的请求、检查和环境会写入工作区,每次调用都会记录到审计文件。除非关闭遥测,否则会发送匿名使用计数。

安装

在 SourceWeft 中

  1. 打开 控制台中的 SignOff Studio,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

SignOff Studio MCP server

Lets AI coding agents prove an API works and that the database holds the right data, on your machine. Works with Claude Code, Cursor, VS Code (Copilot agent mode), Gemini CLI, Claude Desktop and any other MCP client.

bash
npx -y @signoff-studio/mcp init      # in your project: registers the server with the agents it finds
  • Call APIs without curl. http_request returns status, timing, the headers that matter and a compact body (large JSON comes back as its shape plus the first items), with inline checks for status, latency, headers, body text and JSON path.
  • Check the database, read-only. db_query and db_assert run on PostgreSQL, MySQL/MariaDB, SQL Server, Oracle and SQLite. Read-only is enforced by the driver (a read-only transaction or connection), whatever the SQL says.
  • API and database in one step. verify_api_effect answers "POST /orders returned 201, and did the row land?" with one verdict.
  • Checks that persist. save_request and save_db_check write into the project's SignOff Studio workspace, so the next agent session and your teammates rerun them from git.

Runs locally over stdio. No account, no cloud relay. Secrets are redacted by value in every output.

This repository holds the public package docs, registry metadata, example configs and the Claude Code plugin. The server is built from the SignOff Studio product and published to npm as @signoff-studio/mcp.

Install

Needs Node.js 22 or newer (or the SignOff Studio desktop app, which ships the same server and needs no Node).

init only adds a signoff-studio entry: it keeps every other server and setting, never replaces an existing entry and lists each file before writing. Options: --dry-run, --agent claude,cursor,vscode,gemini|all, --global (user-wide), --agents-md (adds a short how-to section to AGENTS.md).

Or register it yourself:

AgentCommand or config
Claude Codeclaude mcp add signoff-studio -- npx -y @signoff-studio/mcp (native Windows: -- cmd /c npx -y @signoff-studio/mcp)
Claude Code pluginclaude plugin marketplace add samuccaya/signoff-studio-mcp then claude plugin install signoff-studio@signoff-studio (adds a verification skill and /signoff-studio:verify)
Cursor.cursor/mcp.json, see examples/cursor
VS Code.vscode/mcp.json, see examples/vscode. The SignOff Studio extension registers the server automatically.
Gemini CLIgemini mcp add signoff-studio npx -- -y @signoff-studio/mcp (native Windows: gemini mcp add signoff-studio cmd -- /c npx -y @signoff-studio/mcp), or examples/gemini
Desktop app (no Node)"<install>\QaSignOff.exe" --mcp-stdio; Microsoft Store install: signoff-mcp.exe. Or use Connect AI agents in the app.

Tools

ToolWhat it does
workspace_statusWhat the server sees: workspace, collections, environment names, DB connections, policy.
http_requestAd-hoc HTTP request with optional inline checks. Works without a workspace.
find_requests, get_requestSearch saved requests by name, "POST /orders", URL or folder; read one.
run_request, run_collectionRun saved requests or a whole collection with their checks; failures show expected vs actual.
get_resultPage through a large result by resultId.
db_connections, db_schemaWorkspace connections (no secrets) and a compact one-line-per-table schema.
db_queryOne read-only SELECT: grid, row count, truncation flag.
db_assertQuery plus expectation (rowCount, scalar, column contains, empty): pass or fail with the actual value.
db_executeWrites, only on connections the owner marked for agent writes, never on production; dry-run (rollback) by default.
verify_api_effectA request plus database checks bound to its response; one verdict.
save_request, save_db_check, create_environmentPersist what the agent verified into the workspace (no overwrite unless asked).
import_requestsImport OpenAPI, Postman, Insomnia, Bruno or cURL.
mock_serverStart or stop a local mock API from a collection or spec.
export_evidenceHTML/Markdown report plus JUnit, and a short summary for the PR.
send_feedbackSend the SignOff Studio team a short note about the tool itself (bug, missing feature, idea); returns a known workaround if there is one. Never URLs, data or code. Off with SIGNOFF_AGENT_FEEDBACK=0.

Read tools carry readOnlyHint, so clients can approve them once. Response bodies and table rows are returned as data, never as instructions.

Safety

  • SQL is read-only unless the owner ticks Allow AI agent writes on a connection. Production-tagged connections never accept agent writes.
  • Database passwords come from the connection's password env key (workspace .env or the environment) or, for vault-only passwords, from the running SignOff Studio app. The model never sees them.
  • Every call is logged to .signoff/agent-audit.jsonl in the workspace (no secrets, no bodies).

Workspace discovery

The agent's open folder (MCP roots), then walking up from the working folder for signoff-workspace.json, then SIGNOFF_WORKSPACE, then the workspace last open in the app. --workspace <dir> overrides. Without a workspace, http_request still works.

Telemetry

Anonymous counts (which tools ran, error codes, which agent) help us fix confusing errors. No URLs, SQL, bodies or names are sent. Turn it off with SIGNOFF_TELEMETRY=0 or DO_NOT_TRACK=1. Details: https://testapilocally.com/mcp#privacy

Privacy Policy

Full policy: https://qasignoff.com/privacy and https://testapilocally.com/mcp#privacy. In short:

  • Collected:
    • Anonymous usage counts: which agent connected, how the server was started, and each tool's outcome or error code, at most once a day per combination.
    • When agent feedback is on, the short reports an agent writes about SignOff Studio itself (a bug, a missing feature, an idea), plus the product version, OS and the names and outcomes of recent tool calls. Reports that contain URLs, hostnames, emails, tokens, SQL, code, file paths or workspace secrets are refused before sending. Never request or response bodies, database rows, credentials or conversation content.
  • Use and storage: to fix bugs and decide what to build. Counts and reports are stored by the usage service (Cloudflare D1); reports are also filed in a private GitHub repository the SignOff Studio team triages. Not shared with or sold to third parties.
  • Retention: usage counts and feedback reports are deleted automatically after 18 months (a daily cleanup in the usage service).
  • Your copy: every feedback report is listed in ~/.signoff-studio/feedback/sent.jsonl.
  • Turning it off: SIGNOFF_AGENT_FEEDBACK=0 turns off feedback. SIGNOFF_TELEMETRY=0 or DO_NOT_TRACK=1 turns off both counts and feedback. The SignOff Studio app's Preferences has the same switches.
  • Contact: via https://qasignoff.com.

Links

© Samuccaya Tech LLP. Free to use under the license in LICENSE.txt.

来源:README.md,提交 f70f66c

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

2
  1. v26.8.42最新Nov 19, 2026
  2. v26.8.41Nov 19, 2026

评分

已满足适用检查项的 40%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分

    请在仓库或注册中心条目中声明许可证。