ppk2-mcp
Enables MCP clients to list, inspect, probe, capture, and analyze current measurements from the Nordic Power Profiler Kit II, returning aggregates and file paths rather than raw samples.
README
ppk2-mcp
A headless driver, command line and MCP server for the Nordic Power Profiler Kit II, wrapped around the device layer of Nordic's own nRF Connect Power Profiler app.
Why it wraps the official app instead of reimplementing it
The PPK2's accuracy does not live in its command set, which is a handful of opcodes. It lives in the per-unit calibration the instrument reports at connect time and in the conversion that turns a 14-bit ADC reading, a measurement range and that calibration into a current - including the filter that suppresses the spike a range switch produces. Those are the parts that change when Nordic ships new instrument firmware, and they are the parts a reimplementation gets wrong quietly: the numbers still look like currents.
So vendor/pc-nrfconnect-ppk is a submodule of the app itself, and
src/device/serialDevice.ts inside it is imported and used as-is.
Nothing in this repository converts an ADC reading. Tracking a new
firmware release is a submodule bump.
What this repository adds is the part the app cannot give a script: the app is an Electron program with a chart, and what a measurement needs is a number attributed to a state.
What it measures that an average does not
The instrument samples current and eight digital channels together at 100 kSa/s. Wire a digital channel to a pin the firmware drives to mark a state - a radio powered down, a peripheral clocked, a task running - and the capture can say what that state cost, rather than what the board averaged while it was happening.
That is the whole design. Captures are stored as RUNS: maximal spans over which the digital word did not change, each with its sample count, mean, extremes and charge. A run file is three orders of magnitude smaller than the samples behind it, loses nothing any per-state question needs, and is small enough to commit beside the result it supports.
Two things are deliberately not smoothed over:
- A sample the instrument's own payload counter says was lost is counted as lost and does not close a run. A dropout is not a state change, and it is not a current either. A capture that lost more than the caller allows is REFUSED rather than averaged, because loss is not spread evenly across the states.
- A state is the MASKED digital word, so channels nothing is clipped to
cannot split a run in two.
probereports every channel so the mask can be chosen from what is actually wired.
Install
Node 20 or newer, and a PPK2 on USB.
git clone --recurse-submodules https://github.com/schuschu/ppk2-mcp
cd ppk2-mcp
npm install
If node is not on PATH, export PPK2_NODE_BIN=/path/to/node/bin and
the ./ppk2 launcher will find it.
On Linux the user needs read and write on the instrument's serial port,
which usually means membership of dialout.
Command line
./ppk2 list
./ppk2 info
./ppk2 probe --seconds 5
./ppk2 capture --out /tmp/run --seconds 60 --mask 0x0f \
--names txrx,addr,txen,rxen --label "what this run is"
probe says which digital channels are wired. capture writes
<out>.runs.csv and <out>.meta.json and prints one line per digital
state. It exits nonzero when the instrument lost more than
--max-dropped-ppm of the stream.
MCP server
./node_modules/.bin/tsx src/mcp.ts
or as a stdio server entry:
{
"command": "npm",
"args": ["--prefix", "/path/to/ppk2-mcp", "run", "--silent", "mcp"]
}
Tools: ppk2_list, ppk2_info, ppk2_probe, ppk2_capture,
ppk2_analyze.
Every tool answers with aggregates and file paths. None of them returns
samples: at 100 kSa/s a minute of capture is six million of them, which
is neither a reply a caller can hold nor one a reader can check.
ppk2_analyze reads a run file back, and given a digital state and a
set of duration edges it splits that state's runs by how long they
lasted - which is how two states one pin cannot distinguish are told
apart by the length of the interval instead.
The instrument is opened per call and closed after it, because one process at a time can hold the port and a server that kept it would lock out both the app and the command line.
Modes
Nothing here ever enables a source output. The instrument is used in whatever mode it is already in, and the harness this was written for is an ammeter in series with a board powered by its own supply - on a Nordic DK, the current measurement header with its jumper removed.
Licence
The vendored app and this wrapper are under Nordic's licence; see
vendor/pc-nrfconnect-ppk/LICENSE.
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。