Spec Driven Development
MCP Server that facilitates spec-driven development workflows, not just Vibe Coding.
作者提供的说明(英文原文)
Full-featured ADB MCP server — 209 tools across 49 modules, from UI to baseband.
这三项永远不会合并成一个分数:意见、测量和人气是不同的东西。
我们还没有为它写通俗的介绍。下面是我们从注册中心读取到的事实。
选择你使用的 AI 应用,然后复制设置。
你需要
你用的是哪个应用?
在上方选择你的应用。我们会记住它,下次按钮会显示“安装到 …”。
打开 Claude Desktop → Settings → Developer → Edit Config,把下面的内容粘贴到 claude_desktop_config.json
{
"mcpServers": {
"deepadb": {
"command": "npx",
"args": [
"-y",
"deepadb"
]
}
}
}在终端中运行这条命令
claude mcp add deepadb -- npx -y deepadb粘贴到项目中的 .cursor/mcp.json,或 ~/.cursor/mcp.json
{
"mcpServers": {
"deepadb": {
"command": "npx",
"args": [
"-y",
"deepadb"
]
}
}
}浏览器会先询问是否打开应用,然后应用会显示设置并请你确认。密钥的值会留空,由你自己填写。
粘贴到项目中的 .vscode/mcp.json
{
"servers": {
"deepadb": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"deepadb"
]
}
}
}浏览器会先询问是否打开应用,然后应用会显示设置并请你确认。密钥的值会留空,由你自己填写。
这个服务器在你的电脑上运行,而 ChatGPT 只能连接互联网上的服务器。请选择其他应用。
你的选择通过 Cookie 保存在这台设备上,无需账户。
这些设置是根据作者在注册中心的条目生成的。如果 README 的写法不同,请以 README 为准。
它在你的电脑上运行;安装后,工具列表会显示在你的 AI 应用中。
我们检查的是任何人都能看到的信息;我们没有阅读代码。不适用的检查项不计入总数。 · 检查于 2026年11月8日
注册名称已通过 GitHub 账户 fullread 验证,该账户拥有此仓库
我们还没有从 GitHub 读取这个仓库
我们还没有从 GitHub 读取这个仓库
它在你的电脑上运行,没有互联网端点
在你的电脑上运行,改用本地设置
它在你的电脑上运行;我们只能读取互联网端点的工具列表
与注册中心里这个分类中所有活跃的服务器比较。只有当发布者拥有该仓库时,星标数才计入。
它在 1 项适用的基础检查中通过了 1 项。我们没有阅读代码。在添加密钥或令牌之前,请先看看这个服务器要求哪些权限,并使用只具备所需权限的密钥。
注册中心没有列出必需的密钥,不过有些服务器会在首次使用时要求你登录。
它在你的电脑上运行。你的 AI 应用会在需要时启动它,数据不会经过作者的服务器(除非它自己调用在线服务)。
GitHub 仓库最后更新于–,最新版本(1.1.6)于6天前进入注册中心。
支持 MCP 的应用,例如 Claude Desktop、Claude Code、Cursor 和 VS Code。你的电脑上还需要有 Node.js。
同一分类中有:Spec Driven Development, Remote Desktop Commander, Tethered Trust
同一分类中其他作者的服务器,按星标数排序。
如何看卡片:★ GitHub 星标i三大支柱:评价 · 检查 · 使用情况i更新 = 最新版本发布到注册中心的时间i
MCP Server that facilitates spec-driven development workflows, not just Vibe Coding.
Remote Desktop Commander: use files, terminal and processes on your own computers from any AI app.
Fit, compatibility and recall checks for AI agents before they buy: verdicts, evidence, receipts.
按原文从作者的仓库复制。PopNx MCP 没有审查或背书其内容。HTML、图片和脚本已被移除。链接会离开本站。
MCP (Model Context Protocol) server for AI agents that need to inspect, control, test, and diagnose Android devices through ADB—or directly on-device through Termux.
209 tools, 5 resources, and 4 prompts across 49 modules—from UI automation and application management to logs, network capture, baseband, firmware, security, and hardware diagnostics.
DeepADB runs as a local MCP server. It uses ADB on a host computer and automatically switches to direct execution when running on Android (detected through /system/build.prop).
git clone https://github.com/fullread/DeepADB.git
cd DeepADB npm ci npm run build npm start
Add build/index.js to your MCP client configuration, then begin with these workflows:
| Goal | Start with |
|---|---|
| Confirm the toolchain and connected device | adb_health_check, then adb_devices |
| Inspect the current Android UI | adb_screen_state, then adb_ui_dump |
| Diagnose a device or app | adb_device_info, adb_logcat, and adb_dumpsys |
For a guided local connection, see MCP Client Configuration. Read SECURITY.md before exposing HTTP, WebSocket, or GraphQL transports beyond loopback.
Using llama.cpp or another local model? See Local LLM clients for direct stdio, authenticated browser HTTP, and small tool profiles. WebSocket is not required.
┌──────────────────────────────────────────────────┐
│ MCP Client │ │ (Any MCP-capable client) │ └──────────────────────┬───────────────────────────┘ │ stdio (JSON-RPC), HTTP (/mcp or /sse), or WebSocket ┌──────────────────────▼───────────────────────────┐ │ DeepADB Server │ │ │ │ ┌─────────────────────────────────────────────┐ │ │ │ Tool Modules (49) │ │ │ │ device │ shell │ packages │ files │ logs │ │ │ │ diagnostics │ ui │ build │ health │ │ │ │ wireless │ control │ logcat-watch │ │ │ │ forwarding │ screen-record │ emulator │ │ │ │ testing │ multi-device │ snapshot │ │ │ │ network-capture │ ci │ plugins │ baseband │ │ │ │ accessibility │ regression │ │ │ │ device-farm │ registry │ at-commands │ │ │ │ screenshot-diff │ workflow │ sensors │ │ │ │ split-apk │ mirroring │ test-gen │ │ │ │ ota-monitor │ ril-intercept │ │ │ │ device-profiles │ firmware-analysis │ │ │ │ workflow-market │ selinux-audit │ │ │ │ thermal-power │ network-discovery │ │ │ │ input-gestures │ wireless-firmware │ │ │ │ app-network │ database-inspector │ │ │ │ runtime-audit │ wear │ │ │ ├─────────────────────────────────────────────┤ │ │ │ Resources (5) │ Prompts (4) │ │ │ └───────────────────┬─────────────────────────┘ │ │ │ │ │ ┌───────────────────▼─────────────────────────┐ │ │ │ ToolContext (unified DI) │ │ │ │ server │ bridge │ deviceManager │ │ │ │ logger │ security │ config │ │ │ └───────────────────┬─────────────────────────┘ │ │ │ │ │ ┌───────────────────▼─────────────────────────┐ │ │ │ Middleware Layer │ │ │ │ OutputProcessor │ SecurityMiddleware │ │ │ │ InputSanitizer │ Logger (stderr-safe) │ │ │ └───────────────────┬─────────────────────────┘ │ │ │ │ │ ┌───────────────────▼─────────────────────────┐ │ │ │ Bridge Layer (auto-detect) │ │ │ │ │ │ │ │ ┌─────────────┐ ┌────────────────────┐ │ │ │ │ │ ADB Bridge │ OR │ Local Bridge │ │ │ │ │ │ (PC mode) │ │ (on-device mode) │ │ │ │ │ │ via adb.exe │ │ via sh/su direct │ │ │ │ │ └──────┬───────┘ └────────┬───────────┘ │ │ │ │ │ │ │ │ │ │ Retry │ Timeout │ Cache │ Serial routing │ │ │ └─────────┼─────────────────────┼─────────────┘ │ └────────────┼─────────────────────┼───────────────┘ │ │ ┌───────▼───────┐ ┌───────▼───────┐ │ ADB Binary │ │ sh / su │ │ (USB/WiFi) │ │ (local) │ └───────┬───────┘ └───────┬───────┘ │ │ ┌─────▼────────────────────▼─────┐ │ Android Device │ └────────────────────────────────┘
DeepADB operates in two modes, auto-detected at startup:
AI Agent (PC) ←→ MCP ←→ DeepADB (PC) ←→ ADB (USB) ←→ Android Device
Standard mode: DeepADB runs on a PC/Mac/Linux host and communicates with the device over USB via ADB. The complete 209-tool surface is registered through the ADB bridge with automatic retry on transient failures; environment-specific tools report their capability requirements when a host cannot run them.
AI Agent (Termux) ←→ MCP (stdio/HTTP) ←→ DeepADB (Termux) ←→ sh/su (local)
When DeepADB runs directly on the Android device (e.g., inside Termux), it auto-detects the environment and switches to LocalBridge. Commands execute directly via sh/su — no ADB server, no USB, no serialization overhead. The same 209-tool MCP surface is registered, with capability-aware guidance when an optional on-device binary such as sqlite3 is unavailable.
1.1.6 device validation (November 1, 2026): the released build's full on-device run passed 836 checks, 0 failed, 3 skipped (839 total) across all 24 suites on the rooted Pixel 6a / Android 16, using Termux Node 24.22.1 and SQLite 3.54.1. The three skips reuse the cached Alpine kernel, initrd and disk; the actual VM boot and shutdown checks passed. Coverage includes the 28 runtime-audit regressions, device SQLite, live Shannon AT/OK, client/plugin compatibility, POSIX security cases and PIN-based UI recovery. The production installation passed eight strict checks before and after promotion, plus eight real Chromium checks and six authenticated WebSocket/GraphQL checks. Following the final documentation audit and verified redeployment, 1.1.6 is the only DeepADB installation retained on the validation Pixel; older installations and rollback copies were removed at the operator's request. Real phone-to-watch validation remains deferred.
The earlier 1.1.6 host/ADB baseline on Windows/Node 24.22.2 passed 783 checks, 0 failed, 18 expected skips (801 total). After the oversized-header test was split into a portable regression, Node 22.23.4 device-free coverage passed 351 checks, 0 failed, 17 expected skips (368 total). Host skips cover device-only or POSIX behavior, unavailable tcpdump and Windows file-symlink privileges; they are distinct from the on-device results above.
v1.1.5 final audit-fix verification (Windows host): the ten device-free suites passed 295 checks, zero failures, and 17 expected skips (312 total) under Node 22.23.4 with coverage. The new hardening suite separately passed 41 checks on Node 24.22.2, with one Windows file-symlink capability skip; 16 other skips in the full CI set require hardware. Eight real-browser CORS/reload checks passed. The Pixel passed eight disposable POSIX path/link/race checks and five focused hardware smoke checks; the latest snapshot code also completed a real schema read through the isolated worker. These initial checks preceded the separate on-device release validation described below.
Final hardening hardware regression (October 31, 2026): all 22 host/ADB suites passed on the Pixel 6a / Android 16 using Node 24.22.2: 728 passed, 0 failed, 18 expected skips (746 total). This includes 41 audit-hardening, 33 local-LLM, 36 SDK/client/plugin compatibility, 93 security and 46 UI checks, plus live SQLite and Shannon validation. The skips are five on-device-only QEMU checks, twelve Windows/POSIX/tcpdump checks and one Windows file-symlink capability case. That last case is additionally covered by the disposable Pixel POSIX probes above. The suite exited successfully, left no ADB forwards, and made no repository file changes. This validates the host build; on-device LocalBridge execution is verified separately during release preparation.
1.1.5 release validation: the production-only package initializes with 209 tools, five resources/templates and four prompts. The staged Pixel build, using Termux Node 24.22.1 and SQLite 3.54.1, passed real SQLite schema/query reads, write rejection, runtime/health checks and a live Shannon handshake. Eight real-browser checks also passed against that on-device process, including authenticated SSE/Streamable HTTP, denied Origins and repeated page reloads. Six optional WebSocket/GraphQL checks passed with their production packages installed. The verified build was promoted to the normal installation path; Android auto-detection and all eight strict device checks passed again there. At that release's validation, the 1.1.4 rollback copy also started on the new runtime. The final host hardware rerun passed 728 checks with zero failures and 18 skips; device-free Node 24 coverage passed 295 checks with zero failures and 17 skips. GitHub CI covers Node 22/24/26 builds and regressions, real-browser MCP compatibility, coverage and dependency auditing, with separate CodeQL analysis for application code and GitHub Actions.
Earlier unreleased host hardware validation (October 31, 2026, before the final seven-finding hardening pass): the expanded suite, including SDK migration, browser reconnect, request-ID/cancellation, plugin callback lifetime, and profile regressions, completed on Pixel 6a / Android 16 with 687 passed, 0 failed, and 17 expected skips (704 total). Live llama.cpp checks separately exercised stdio, authenticated browser Streamable HTTP, and legacy SSE with the inspect profile; see LOCAL_LLMS.md for the tested model and browser HTTP reload recovery and its limits. Phone-to-watch Wear validation is deferred.
v1.1.4 host hardware validation: the full Pixel 6a / Android 16 regression completed with 604 passed, 0 failed, and 17 expected environment-specific skips (621 total). This includes live AT/OK validation on /dev/umts_router, a guarded read-only SQLite schema snapshot with cleanup, the unified runtime audit, per-app route context, and the single-device Wear preflight. The skips are the five on-device-only QEMU checks plus twelve Windows-host or unavailable-tcpdump security checks.
v1.1.3 hardware baseline: validated on a Pixel 6a (Android 16, Termux + Magisk + QEMU 10.2.1) across a four-cell test matrix — host (ADB) and on-device (LocalBridge), each with and without a device PIN — with 0 failures in every cell. The v1.1.3 suite contained 577 tests in the host/ADB configuration and 613 on-device; the difference is the QEMU virtualization and Alpine VM-boot suites, which run only on-device:
只显示了开头部分;其余内容请在仓库中阅读