Plctap vs Predictive Maintenance MCP Server
A side-by-side look at the Plctap and Predictive Maintenance MCP Server MCP servers: transport, auth, tools, quality signals, and the config to install either one in Claude, Cursor, Windsurf or VS Code.
Verdict: choose Plctap for industrial & iot work as a local process; choose Predictive Maintenance MCP Server for industrial & iot as a local process. You can also run both side by side in one mcpServers block.
Which MCP server should you choose?
Choose Plctap when
- You need dedicated capabilities in the Industrial & IoT domain.
- You prefer a local stdio subprocess transport.
- You want tools like
list_protocols, probe_device, plc_read.
Choose Predictive Maintenance MCP Server when
- You need dedicated capabilities in the Industrial & IoT domain.
- You prefer a local stdio subprocess transport.
- You have the keys it needs:
PDM_PROJECT_DIR. - You want tools like
list_signals, generate_test_signal, load_signal.
Feature & specification comparison
Rows where the two servers differ are marked.
| Spec | ||
|---|---|---|
| Category | Industrial & IoT | Industrial & IoT |
| Quality signal | 56/100 · Good | 63/100 · Good |
| Install command | Confirmed | Confirmed |
| Transport | Local subprocess (stdio) | Local subprocess (stdio) |
| Auth | Not declared | Not declared |
| Pricing | Not declared | Not declared |
| Required env vars | None listed | PDM_PROJECT_DIR |
| Verified owner | Community listing | Community listing |
| GitHub stars | 2 | 106 |
| Directory activity | 3 views 0 installs 0 upvotes | 1 views 0 installs 0 upvotes |
| Tools | 15 | 34 |
| Full listing | View Plctap | View Predictive Maintenance MCP Server |
Tools & capabilities
Plctap tools 15
list_protocols列出本 server 支持的工业协议及其能力。 返回每个协议的端口提示、地址模型、数据类型和品牌线索, Agent 可据此推断 "这个设备该用什么协议" 而无需问用户。 不确定协议名时先调用本工具。 不确定协议/端口时先调用 detect_device。probe_device测试能否连上 PLC 并通信, 失败时给出层级归因。 reachable = 传输层可达 (TCP 已建立; 设备回异常响应也算在线)。 failure_class 取: - connection_refused: 端口没人监听/网络不可达 (查网络与端口) - timeout: 连接超时 (查网络路由/防火墙) - connected_but_no_reply: TCP 通了但设备不回话 (查协议配置) - exception_response: 设备回异常码 (查 unit/寄存器配置, 带 exception_code) 典型用法: 设备"读不到数据"时先调本工具分层定位, 再决定下一步。plc_read从 PLC 读取数据区并按数据类型解释。 address/count 为通用地址与数量, 语义由协议决定: - Modbus: address 为 0 基寄存器地址, count 为寄存器个数 - FINS: address 为字地址, count 为字数 - MELSEC: address 为起始编号, count 为点数 (位软元件按 16 点/字) - S7: address 为字节地址, count 为**字节数** (count=4 + uint16 → 2 个值) - EtherNet/IP: address 为 tag 名字符串 (如 "alpha[0]"), count 为元素个数 - OPC UA: address 为 NodeId 字符串 (如 "ns=2;i=5" / "ns=2;s=Demo.Double"), count 为数组节点返回元素上限 (0=全部); 值按 UA 内建类型原生返回 datatype 取 uint16/int16/float32, None 返回原始 16 位值 + 所有常见数据类型的多解释 (interpretations 字段), 便于 Agent 识别正确的数据类型。 byteorder 仅影响 float32 寄存器对顺序 (big=ABCD, little=DCBA)。 options: 协议特有参数, 由各协议 adapter 自行定义与校验。 用 list_protocols 查看每个协议的 read_options 说明。plc_browse浏览地址空间: 从 node 展开一层子节点 (诊断场景的"列目录")。 返回 {node, children, total, shown, truncated}: children 每项含 node_id / display_name / node_class; 超出输出预算时 truncated=true 且 total 给出全量数 —— 需要更深层级就对着子 node_id 再调一次。 默认从 Objects 文件夹 (ns=0;i=85) 起步; 对 plc_read 前先摸清 设备地址空间结构时用。 当前仅 OPC UA 支持 (会话协议无帧概念, browse 是它的"读目录"形态); 其他协议用 list_protocols 的 read_options 直接构造地址。 输出预算: 单次最多 200 个子节点 (约 30KB), 防止大地址空间撑爆上下文。diagnose综合观测给出结构化的故障候选结论 (确定性规则, 不编故事; 全程只读)。 三种证据可任意组合 (至少给一种): - frame_hex: 一帧报文 hex (解析 + 校验 + 规则匹配) - log_snippet: 通信日志文本, 自动提取其中的 hex 帧逐帧解析 - host + port: 连上设备做一次探测, 把探测归因纳入推理 返回 DiagnosticReport: candidates 按 confidence 降序, 每条含 symptom/root_cause/evidence(证据链)/suggested_action/next_tools; evidence 与 observations 保留原始观测供复核。空 candidates = 知识库未覆盖, 不硬凑结论。parse_frame把一帧报文 hex 逐字段结构化解析 (不需要连接设备)。 每个字段带 byte_offset 与 raw_hex 证据; 支持正常请求/响应与异常 响应帧; 畸形帧不抛错, 而是尽量解析并在 errors 里说明哪里坏 —— "解析失败的方式"本身是诊断证据。 direction: "auto" 按帧结构判别, 也可显式传 "req" / "resp"。 frame_format (仅 melsec): "3e_binary" / "3e_ascii" / "4e_binary" / "4e_ascii"。 - modbus: 异常帧/非 12 字节按响应, 12 字节按请求 (fc05/06 的响应 与请求逐字节相同, 任一解释一致) - fins: TCP 命令 0x00 按请求 / 0x01、0x02 按响应; 0x04 按 FINS ICF bit6 判别 (响应置位) - melsec: 副头部 50 00/54 00 按请求, D0 00/D4 00 按响应 (ASCII 帧 "5000"/"5400"/"D000"/"D400" 同理) - s7: rosctr=0x01 (Job) 按请求, 0x03 (Ack_Data) 按响应; 显式传 direction 时按传入值validate_frame对一帧报文跑规范校验清单, 逐项 pass/fail (不需要连接设备)。 Modbus TCP: MBAP 长度一致性、协议号、功能码、unit 范围、PDU 自洽、数量/地址边界、异常码 (Modbus TCP 无 CRC, RTU 才有)。 FINS/TCP: magic、TCP 长度自洽、TCP 命令合法、error 字段、 FINS 端结码、载荷完整性。 MELSEC 3E: 副头部、数据长度自洽、命令/软元件代码、结束代码。 direction 取 req 或 resp。parse_pcap解析 Wireshark 导出 pcap: 按 TCP 流聚合载荷 -> 按协议切帧 -> 逐帧 parse_auto。 每个流返回完整帧序列 (带结构化解析) 与尾部半帧 (partial, 截断也是 诊断信息)。protocol 缺省时按"完整帧数最多的协议"自动判别。 需要可选依赖 scapy (uv sync --extra eval)。大批量帧场景比逐条 frame_hex 高效得多。 输出受 64KB token 预算约束: 超出时按流序/帧序装帧并截断, 末尾追加 一条 flow 以 "truncated:" 开头的 sentinel 流 (frames 为空, 带 total/shown 计数)。需要其余帧时用 protocol= 过滤或拆分 pcap 再解析。start_listener起钓鱼模式监听: 待测设备只能当 client 时, 立假 server 钓出它的帧行为。 mode 三档: record_only=只收帧不回复 (纯被动) / respond_normal=对读类 请求回最小"正常响应" (数据恒 0, 目的只是让设备继续吐帧, 非通用模拟器) / inject_errors=正常回帧但按 faults 列表轮转注入确定性故障 (评测语料 生产 + 诊断引擎回归)。 faults (仅 inject_errors): modbus 可选 exception/bad_length/truncate, fins/melsec 可选 end_code/bad_length, 全协议通用 garbage。 收下的帧用 get_listener_frames 取, 再喂 parse_frame/diagnose。 port=0 由系统分配, 返回实际端口。建议收满样本后 stop_listener, 并提醒用户恢复设备原配置。stop_listener停掉指定端口的监听, 返回收/发帧统计。get_listener_frames取监听收下的帧 (direction/peer/frame_hex), 供 parse_frame/diagnose 分析。 取最新 limit 条; 环形缓冲硬上限 1000 条 (limit 超限自动截到缓冲 大小, limit<=0 返回空) —— 输出量受 limit 与缓冲上限双重约束。start_proxy起透明代理: 上位机 → 代理 → 真实 PLC, 透传同时按协议分帧录制双向帧。 现场联调时把上位机目标地址改成本代理, 无需 Wireshark 即可拿到全部 交互帧 (get_proxy_frames), 再喂 parse_frame/diagnose 做在线分析。 protocol 当前支持 modbus/fins/melsec (S7 TPKT 分帧暂不支持); listen_port=0 由系统分配。代理是诊断设施: 只透传与录制, 不改写帧。- +3 more on the listing page
Predictive Maintenance MCP Server tools 34
list_signalsList signal files on disk or signals loaded in the repository. scope='disk' (default): files under data/signals/ that load_signal can open — use before loading. scope='memory': signals currently cached in the in-memory repository with their metadata (signal_id, sampling_rate, declared unit) — use to see which signal_ids are available for analysis. Args: ctx: MCP context. Unused — see this module's docstring on logging. scope: 'disk' for loadable files, 'memory' for loaded signal_ids. Returns: Dict with scope, count, and either 'files' (relative paths, disk) or 'signals' (StoredSignalInfo entries, memory).generate_test_signalGenerate a synthetic test signal, save it, and load it into the repository. The signal is written to data/signals/ with a timestamped filename and a companion _metadata.json declaring sampling_rate and signal_unit='g' (synthetic acceleration), then auto-registered in the repository — the returned signal_id is immediately usable by every analysis, diagnosis, and ISO severity tool with no manual steps. Signal content: 'bearing_fault' = 10 Hz impacts modulating a 1 kHz carrier; 'gear_fault' = 200 Hz mesh tone + harmonics; 'imbalance' = 25 Hz (1500 RPM) tone; 'normal' = broadband noise. Args: signal_type: Synthetic fault pattern to generate. duration: Signal duration in seconds (10 s gives 0.1 Hz resolution). sampling_rate: Sampling frequency in Hz. noise_level: Additive white-noise amplitude. random_seed: Seed for reproducible noise (None = non-deterministic). ctx: MCP context. Unused — see this module's docstring on logging. Returns: StoredSignalInfo of the auto-loaded signal (signal_id, declared sampling_rate and unit 'g').load_signalLoad one signal — or a batch — into the in-memory repository. Once loaded, reference the signal by its signal_id in every analysis, diagnosis, report, and prognostics tool (the load -> analyze -> diagnose -> report flow uses signal_id as the single handle). Batch form: pass a LIST of file paths (e.g. for training sets). The batch is fail-fast and atomic — all paths and derived ids are validated up front, and on the first problem ONE error names the offending entries and nothing is loaded. One declared sampling_rate/ signal_unit applies to all files; per-file metadata wins only when the parameter is omitted. Custom signal_id is not allowed for a batch (ids derive from each file's relative path). signal_id default: the path relative to data/signals/ with separators replaced by underscores — 'real_train/baseline_1.csv' loads as 'real_train_baseline_1', so same-named files in different folders never collide silently. Re-loading a path whose id already exists is an explicit error unless overwrite=True. Signal unit discipline: ISO 20816-3 severity verdicts require a DECLARED unit — either via this parameter or a 'signal_unit' field in the companion _metadata.json (explicit parameter wins). Units are never guessed from signal amplitude; without a declared unit the ISO severity block is refused with a structured reason and remedy. Raw binary files (.bin/.raw/.dat): a headerless raw waveform loads only with a declared decode contract — sample_format AND sampling_rate are REQUIRED, either as explicit parameters here or as fields of the companion <stem>_metadata.json next to the file (explicit parameter wins). The other raw parameters carry documented defaults, applied by the repository after that merge: byte_order 'little', n_channels 1, channel_index 0, header_offset 0, no scale_factor. Integer sample formats (int16/int32) decode to raw ADC counts — declare scale_factor to convert counts into the declared physical unit (there is no implicit normalization). In a batch the raw parameters broadcast to ALL files, exactly like sampling_rate. Declaring raw parameters for a self-describing format (.csv, .npy, ...) is refused as a contradiction. With n_channels > 1 each load extracts ONE channel and the DERIVED id gains a _ch<channel_index> suffix; an explicit signal_id is used verbatim — no suffix applies. Args: ctx: MCP context. Unused — see this module's docstring on logging. filepath: Filename relative to data/signals/ or absolute path — or a list of such paths for an atomic batch load. signal_id: Custom ID (single-file loads only; default derives from the relative path). sampling_rate: Sampling rate in Hz (overrides metadata file). Required for raw binary files (here or in the companion). signal_unit: Declared signal unit — 'g' or 'm/s2' (acceleration), 'mm/s' or 'm/s' (velocity). Overrides the metadata file. overwrite: Replace existing entries on signal_id collision instead of raising. sample_format: Raw files only — declared sample dtype ('float32', 'float64', 'int16', 'int32'). REQUIRED for .bin/.raw/.dat (here or in the companion metadata). byte_order: Raw files only — declared endianness ('little' or 'big'); documented default 'little'. n_channels: Raw files only — interleaved channel count in the file; documented default 1. channel_index: Raw files only — 0-based channel to extract; documented default 0. header_offset: Raw files only — bytes to skip before the first sample; documented default 0. scale_factor: Raw files only — optional multiplier applied after decoding (e.g. ADC counts -> physical unit); default: no scaling. Returns: StoredSignalInfo for a single load; a list of StoredSignalInfo (input order) for a batch. Raw loads record the effective decode parameters under raw_format. Raises: ValueError: If signal_unit is invalid, the signal data cannot be loaded, a signal_id collides without overwrite=True, a batch contains any invalid entry (nothing is loaded), a raw binary file is missing a required declaration (ONE message names everything missing plus both remedies), or raw parameters are declared for a self-describing format.get_signal_infoGet metadata for a stored signal without loading the full array. Includes the COMPLETE companion-metadata dict (source_metadata: rpm/ shaft_speed, reference frequencies, ...) alongside the repository fields (sampling_rate, declared signal_unit, shape, timestamps). Args: ctx: MCP context. Unused — see this module's docstring on logging. signal_id: ID of a signal previously loaded via load_signal. Returns: StoredSignalInfo with source_metadata populated from the companion _metadata.json (empty dict when the file has none). Raises: ValueError: If the signal_id is not in the repository.clear_signalsRemove one signal — or all signals — from the in-memory repository. Args: ctx: MCP context. Unused — see this module's docstring on logging. signal_id: ID to remove; None (default) clears the whole cache. Returns: Dict with cleared_count, plus signal_id and status ('removed' or 'not_found') for single-signal calls.analyze_fftPerform FFT (Fast Fourier Transform) analysis on a stored signal. FFT analysis converts the signal from time domain to frequency domain, allowing identification of harmonic components and faults that manifest at specific frequencies. Requires the signal loaded via load_signal() first; the sampling rate comes from the stored signal metadata. By default analyzes the LEADING 1.0-second segment (deterministic: two identical calls return identical results). Set segment_duration=None to analyze the entire signal, or pass random_seed to sample a seeded random segment position instead. Args: ctx: MCP context. Unused — see this module's docstring on logging. signal_id: ID of the stored signal (from load_signal). max_frequency: Maximum frequency to analyze (default: Nyquist frequency) segment_duration: Duration in seconds to analyze (default: leading 1.0 s). Set to None to analyze the full signal. random_seed: Seed for random segment position (default: None = deterministic leading segment). Returns: FFTResult with top peaks, dominant peak, and spectrum stats. Raises: ValueError: If the signal_id is not loaded, or the stored signal has no sampling rate.analyze_envelopeEnvelope-spectrum analysis of a stored signal (bearing fault screening). THE unified envelope tool: bandpass filter -> Hilbert envelope -> mean subtraction + Hann window -> FFT -> top peaks. The mean subtraction/window step is an intentional U9 fix: the envelope's DC leakage used to bury the low-frequency FTF zone. Requires the signal loaded via load_signal() first; the sampling rate comes from the stored signal metadata. The requested band must fit the signal: an invalid band (low <= 0, low >= high, high > Nyquist) raises a ValueError — it is NEVER silently clamped. The band used is echoed in the result. By default analyzes the LEADING 1.0-second segment (deterministic: two identical calls return identical results). Set segment_duration=None to analyze the entire signal, or pass random_seed to sample a seeded random segment position instead. No reference bearing frequencies are assumed: compare the returned peaks against frequencies computed for the actual bearing and shaft speed (check_bearing_faults or calculate_bearing_characteristic_frequencies). Args: ctx: MCP context. Unused — see this module's docstring on logging. signal_id: ID of the stored signal (from load_signal). filter_low: Bandpass low edge in Hz (default: 500). filter_high: Bandpass high edge in Hz (default: 5000). Must not exceed the signal's Nyquist frequency. num_peaks: Number of top peaks to return (default: 5). segment_duration: Duration in seconds to analyze (default: leading 1.0 s). None analyzes the full signal. random_seed: Seed for random segment position (default: None = deterministic leading segment). Returns: EnvelopeResult with the band actually used, top peaks, and comparison guidance. Raises: ValueError: If the signal_id is not loaded, the stored signal has no sampling rate, or the band is invalid vs Nyquist.analyze_statisticsCalculate statistical parameters of a stored signal for diagnostics. Statistical parameters are key indicators for diagnostics: - RMS: Effective value, correlated to signal energy - Crest Factor: Indicates presence of impulses (high = possible faults) - Kurtosis: Measures impulsiveness (excess kurtosis; >0 = non-Gaussian, >3 = strong impulses) - Peak-to-Peak: Signal range Requires the signal loaded via load_signal() first. Statistical parameters are screening indicators, not definitive diagnostics — combine with frequency-domain evidence. **Signal units:** all values are in the signal's native unit. The unit is reported only when DECLARED — load_signal(signal_unit=...) or the companion _metadata.json — and never guessed from signal amplitude. ISO 20816-3 severity tools refuse to produce a verdict until the unit is declared. Args: signal_id: ID of the stored signal (from load_signal). Returns: StatisticalResult with all statistical parameters Raises: ValueError: If the signal_id is not loaded.extract_features_from_signalExtract time-domain features from a stored signal using sliding windows. Segments the signal into overlapping windows and extracts 17 statistical features from each segment. Features include: mean, std, RMS, kurtosis, crest factor, entropy, etc. Requires the signal loaded via load_signal() first; the sampling rate comes from the stored signal metadata. Returns an in-memory summary only — no CSV is written to data/signals/. Args: signal_id: ID of the stored signal (from load_signal). segment_duration: Duration of each segment in seconds (default: 0.1) overlap_ratio: Overlap between segments, 0-1 (default: 0.5 = 50%) ctx: MCP context. Unused — see this module's docstring on logging. Returns: FeatureExtractionResult with features matrix and metadata Raises: ValueError: If the signal_id is not loaded, or the stored signal has no sampling rate. Example: extract_features_from_signal( "healthy_motor", segment_duration=0.2, overlap_ratio=0.5 )compute_power_spectral_densityCompute Power Spectral Density (Welch method) for a stored signal. Requires signal loaded via load_signal() first. Args: signal_id: ID of the stored signal. nperseg: Samples per FFT segment (default 256). noverlap: Overlap between segments (default 128). window: Window function (default 'hann').compute_spectrogram_stftCompute STFT spectrogram for a stored signal. Returns time-frequency summary (no full 2D array). Use for detecting time-varying frequency content (transient faults, speed changes). Args: signal_id: ID of the stored signal. nperseg: Samples per STFT segment (default 256). noverlap: Overlap between segments (default 128). window: Window function (default 'hann').assess_severityAssess vibration severity (ISO 20816-3 zones A-D) and alert level. THE unified severity tool: ISO zone assessment and alert classification in one call. Zone boundary values are those of ISO 10816-3:2009 (ISO 20816-3:2022 merges zones A/B — provenance is noted in the result). Scope: machines rated above 15 kW; a declared machine_power_kw below 15 kW is refused. Input routes (exactly ONE required): - signal_id: a stored signal (load_signal first). Sampling rate AND declared unit come from the stored metadata; an undeclared unit is refused, never guessed from amplitude. - rms_velocity_mm_s: a direct broadband RMS velocity reading in mm/s (e.g. from a portable instrument) — no unit declaration needed. Args: ctx: MCP context. Unused — see this module's docstring on logging. signal_id: ID of the stored signal (mutually exclusive with rms_velocity_mm_s). rms_velocity_mm_s: Direct broadband RMS velocity in mm/s (mutually exclusive with signal_id). machine_group: 1 (large, >300 kW) or 2 (medium, 15-300 kW). Ignored when custom thresholds are given. Default 2. support_type: 'rigid' or 'flexible'. Ignored when custom thresholds are given. Default 'rigid'. thresholds: Optional custom zone boundaries {'warning': A/B, 'alarm': B/C, 'danger': C/D} in mm/s, strictly increasing — replaces the ISO table for this call. machine_power_kw: Rated machine power, if known. Declared values below 15 kW are refused (out of ISO scope); None means unknown and is not refused. rpm: Operating speed in RPM (signal route only: selects the 2 Hz band lower edge below 600 RPM). Returns: VibrationSeverityResult (status='assessed') with zone, severity, boundaries, derived alert_level/exceeded_threshold, and threshold provenance. Raises: ValueError: On route misuse (both/neither inputs), undeclared signal unit, missing sampling rate, Nyquist below the ISO band, declared power below 15 kW, negative RMS, or invalid custom thresholds.- +22 more on the listing page
Ready-to-paste client configs
Paste either block (or both) into your client config, e.g. claude_desktop_config.json or ~/.cursor/mcp.json. We only show a config when the install command is confirmed.
Plctap
{
"mcpServers": {
"plctap": {
"command": "uvx",
"args": [
"plctap"
]
}
}
}
Predictive Maintenance MCP Server
{
"mcpServers": {
"predictive-maintenance-mcp-server": {
"command": "uvx",
"args": [
"predictive-maintenance-mcp"
],
"env": {
"PDM_PROJECT_DIR": "YOUR_PDM_PROJECT_DIR_HERE"
}
}
}
}
Frequently asked questions
Plctap is categorized under Industrial & IoT and uses a local stdio subprocess. Predictive Maintenance MCP Server belongs to Industrial & IoT and uses a local stdio subprocess. Pick Plctap for industrial & iot work and Predictive Maintenance MCP Server for industrial & iot.
Copy the mcpServers block from the config section on this page into your client's config (for example claude_desktop_config.json or ~/.cursor/mcp.json), then fully restart the client.
Plctap has no declared pricing — check the project's own site. Predictive Maintenance MCP Server has no declared pricing — check the project's own site. If a server needs environment variables such as API keys, define them under the env key in your client's MCP config.
Yes. MCP clients support multiple servers: add plctap and predictive-maintenance-mcp-server as separate keys in the same mcpServers object, and your assistant will call tools from either one.
On AllMCPs, Plctap has a quality signal of 56/100 (Good) with 3 views, 0 installs and 2 GitHub stars. Predictive Maintenance MCP Server has 63/100 (Good) with 1 views, 0 installs and 106 GitHub stars.