{"slug": "testing-claude-for-manufacturing-with-predictive-maintenance-mcp", "title": "Testing Claude for Manufacturing with Predictive Maintenance MCP", "summary": "An independent study by Predictive Maintenance MCP maintainer LGDiMaggio compared Anthropic's Claude for Manufacturing plugin, released 12 November 2026 with seventeen industrial skills, against a dedicated Predictive Maintenance MCP diagnostic toolkit on real laboratory bearing recordings embedded in synthetic maintenance scenarios. The experiment measured diagnostic conclusions, uncertainty, evidence preservation and computational overhead, testing whether a manufacturing agent given standard condition-monitoring tools produces more reproducible, verifiable fault diagnoses than one that generates its own signal-processing code. The work is independent and not sponsored, reviewed or endorsed by Anthropic.", "body_md": "At 10:30, a transfer pump stops after an unusual noise. The technician believes the cause is familiar.\n\nAn outer-race bearing failure, similar to one repaired several months earlier. The maintenance history makes that explanation plausible. But a vibration recording taken twenty minutes before the shutdown points toward a different fault location.\n\nWhat should an AI maintenance assistant do? Follow only the technician’s suggestions, trust its own signal analysis, or keep several explanations open until more evidence is available?\n\nThis is one of the maintenance scenarios constructed for the study presented here. The vibration recording comes from a real laboratory dataset, while the pump, technician observations and maintenance history are synthetic.\n\n**Disclosure.** I maintain [Predictive Maintenance MCP](https://github.com/LGDiMaggio/predictive-maintenance-mcp), one of the two systems examined here. This is independent work, not sponsored, reviewed or endorsed by Anthropic. AI tools were used throughout: the figures were prepared with AI assistance, and AI tools drafted the manuscript from the study files. I defined the research direction, did the experiments and revised the final text.\n\nOn 12 November 2026, Anthropic released [*Claude for Manufacturing*](https://claude.com/blog/claude-for-manufacturing), a plugin containing seventeen skills for industrial workflows [1]. One of them, breakdown-diagnosis, guides an AI assistant through a maintenance investigation, combining equipment history, technician observations, sensor data and proposed corrective actions.\n\nThat workflow intersects with [Predictive Maintenance MCP](https://github.com/LGDiMaggio/predictive-maintenance-mcp) (PMM), an independent open-source toolkit that provides specialized condition-monitoring and diagnostic methods through the Model Context Protocol. Claude for Manufacturing can generate its own signal-processing code, while PMM makes existing analytical and industrial standard procedures available as tools.\n\nThis raises a question that matters beyond the ability to perform a calculation.\n\n*What changes when the same manufacturing agent can use a dedicated diagnostic toolkit instead of constructing the analysis itself?*\n\nFor a maintenance engineer, finding a plausible fault is only part of the problem. The analytical procedure should also be understandable, reproducible and open to independent verification. When the same type of measurement is examined again, engineers need to know which methods and criteria were applied, and how strongly the evidence supports the conclusion.\n\nThe experiment therefore examines not only what Claude identifies, but also what changes when diagnostic calculations and evidence rules are provided by a dedicated software component rather than generated during each investigation.\n\nTo investigate this, I conducted experimental runs using real laboratory bearing recordings embedded in synthetic maintenance scenarios. The comparison examines diagnostic conclusions, uncertainty, evidence preservation and computational overhead.\n\nA skill, in Claude’s terms, is a written procedure: a Markdown file, with reference documents and sometimes scripts, that the model reads and follows when a matching request arrives. Claude for Manufacturing packages seventeen of them, twelve single-purpose skills and five chains that link them, around one dataset joined from a plant’s ERP, maintenance and production exports [2]. Data arrives through connectors for Microsoft 365, Oracle NetSuite, Infor M3 and Autodesk Fusion, or as scheduled file exports. Every skill ends at a draft that a person reviews, and nothing is written to a system of record without approval. The repository describes the plugin as a reference implementation, a starting point for organizations to adapt [1, 2].\n\nThe breakdown-diagnosis skill starts from the premise that, in a breakdown, more time usually goes into working out what is wrong than into the repair. The technician has several sources of evidence and no time to combine them, so the skill has Claude do it in a fixed order.\n\nThe procedure begins with safety. Every diagnosis opens with a reminder to follow the plant’s lockout procedure, and six listed conditions, among them a seized rotating part, a defeated guard and an electrical fault, raise a safety flag that must appear before any diagnostic suggestion. Claude then records the event in the technician’s own words, reads the sensor data from before the failure, and queries the maintenance history of the machine and of its sister machines. The queries run as code and the work orders they return are listed. Close-out notes are grouped by their exact text, so that a sentence repeated on unrelated machines is treated as a template and not cited as evidence. Photographs and the manufacturer’s troubleshooting guide are read where they exist.\n\nThe result is a ranked list of candidate causes. Each carries its evidence, the work orders behind it, the cheapest check that would separate it from the next candidate, and the expected repair. The skill also drafts a change to the preventive-maintenance plan and a downtime record that stays advisory until a technician confirms the diagnosis.\n\nOn sensor data the skill is brief. Claude should read the shape of the data, such as a slow climb or a step, as well as alarm crossings, check that the data cover the event, and say so when there are none, which the skill treats as the normal case. It expects trend exports from a plant historian, the system that archives process signals, and prescribes no signal-processing method. When the input is a raw vibration waveform, the analysis is left to Claude, using general-purpose Python or whatever specialised tool is connected (Figure 1).\n\n[Predictive Maintenance MCP](https://github.com/LGDiMaggio/predictive-maintenance-mcp), PMM from here on, is a server that runs locally and exposes condition-monitoring functions as tools a language model can call [3, 4]. It contains no language model. A tool runs deterministic code on a signal and returns a compact structured result, while the waveform stays on the user’s machine. The main branch lists 44 tools and four guided prompts, organised along the processing blocks of ISO 13374 [8], from signal acquisition to diagnostics, prognostics and decision support.\n\nIts functional areas correspond to questions that arise in a maintenance investigation. Spectral and statistical tools describe a signal before anyone diagnoses it. A severity tool places overall vibration in the zones of ISO 20816–3 [9] and declines to answer when the unit is not declared. An asset ledger records each identified measurement locally, so that a later reading can be compared with a declared healthy baseline, which is how a route technician’s usual question, what has changed on this machine, gets an answer. Anomaly detection can be trained on a machine’s healthy recordings. A remaining-life tool works from repeated measurements and refuses to extrapolate from one. Documentation tools search equipment manuals, and reporting tools produce documents whose evaluative wording is written by the server.\n\nI list these because they describe what the server is, and this study examined only a small part of it. The 71 runs evaluated one function closely, the bearing-fault check, together with one side effect of loading a recording. Statistics, spectra, severity, history look-ups and the change assessment were called in some runs without their results being evaluated. Anomaly detection, prognostics, documentation retrieval and reporting were not exercised, and nothing below is evidence about them. The revision tested is the unreleased main branch and the evidence rules described next are not yet in the published 0.13.0 package.\n\nThe two projects address different aspects of industrial maintenance. Claude for Manufacturing provides workflows for investigating breakdowns, connecting operational information and preparing maintenance decisions. PMM focuses on condition monitoring, offering specialized tools for signal processing, diagnostic assessment, measurement history, anomaly detection and prognostics. Figure 2 summarizes their complementary capabilities and illustrates how these tools can support a broader maintenance investigation when used together. The experiments presented here evaluate only a subset of these possibilities.\n\nThe bearing check rests on a physical regularity. When a rolling bearing has a damaged spot on one of its races, a ball strikes the spot each time it passes. The impacts repeat at a rate fixed by the shaft speed, the bearing’s geometry and the location of the damage, so outer-race and inner-race damage have different characteristic frequencies. The impacts are weak next to the rest of the machine’s vibration, but filtering a high-frequency band and tracing its envelope recovers their rhythm, and the spectrum of that envelope shows each repetition rate as a line.\n\nPMM computes this and then applies a fixed rule (Figure 3). A line counts only if it rises above the local noise floor by a significance threshold. If one fault type is supported, the result may name it, at a claim level the server calls single-spectrum evidence. If several are supported, none is singled out, and the server states that the order of their levels is not a ranking. No outcome is called a confirmed fault, and finding nothing is not reported as health. The server writes the assessment sentence itself and instructs the model to quote it.\n\nNeither project refers to the other. The reason to try them together is the step the skill leaves open, which Claude can fill by writing code or by calling a tool built for the purpose. It was not obvious what difference that choice would make to the investigation.\n\nThe distinction is not simply between two ways of calculating a vibration spectrum. In the Python-only configuration, Claude decided how to process the signal, generated the analysis code and executed it through its available Python environment. With PMM, Claude selected and called specialized tools, while the numerical processing and diagnostic evidence rules were implemented and executed by the local server. The vibration recording remained on the user’s machine, and PMM returned structured results for Claude to interpret.\n\nThis changes where the analytical methods are defined and controlled. Code generated for an individual investigation is not guaranteed to be identical across independent runs, whereas PMM provides predefined implementations that can be tested, reviewed and reused separately from the language model. Neither approach guarantees a correct diagnosis: Claude still decides how to interpret the available evidence and communicate the findings to the maintenance engineer.\n\nThe Model Context Protocol (MCP), a standard through which a model’s host application discovers and calls external tools, made the connection a matter of configuration [5]. With PMM registered as an MCP server in Claude Code, its tools were available in the same session as the plugin’s skills, and no integration logic code was written.\n\nFour set-ups were compared. In the first, Claude for Manufacturing had access to general-purpose Python and nothing else. In the second, PMM was used on its own, with the recording but no maintenance records, as a reference point and not a comparable workflow. In the third configuration, Claude for Manufacturing followed the same breakdown-diagnosis workflow, but Predictive Maintenance MCP was also available as an external tool server. The prompt was identical to the first configuration and did not mention PMM, allowing me to observe whether Claude would discover and use its diagnostic tools on its own. In the fourth configuration, I added five sentences to guide how Claude handled the diagnostic evidence returned by the server. These instructions asked Claude to preserve the original assessment and its level of certainty, report the relevant numerical findings, distinguish measured shaft speed from declared values, and disclose when loading a recording updated the local asset ledger.\n\nAll experiments used fixed software versions. The unmodified Claude for Manufacturing plugin (v1.0.0), Predictive Maintenance MCP at commit 3707552, and Claude Code 2.1.326 running the claude-opus-5-5 model. Each run started in a fresh session with the relevant case files.\n\nI used real vibration recordings from two public laboratory datasets: MFPT [6] and the Paderborn University Bearing DataCenter [7]. To simulate a maintenance investigation, I combined these recordings with synthetic information about pumps, previous work orders, shift logs, and technician observations. This distinction is important, as **the vibration measurements were real, but the industrial scenarios were constructed for the experiments.** In the MFPT study, the laboratory fault labels were withheld until the runs were completed. Even when an agent identified the correct label, this did not amount to confirming the root cause of a real machine failure.\n\nThe study comprised three stages, for a total of 71 runs. I began with 17 exploratory runs across six cases to understand how the systems behaved and identify questions worth investigating further. I then selected three MFPT cases and repeated each five times under three configurations: -\n\nThese 45 runs form the main quantitative comparison presented in the article.\n\nFinally, I performed nine additional runs using three Paderborn bearing specimens. Here, I compared Claude for Manufacturing using Python, PMM used independently, and Claude for Manufacturing with PMM connected, with one run per specimen and configuration. These observations are reported separately from the repeated MFPT experiments.\n\nThe results were evaluated using automated checks for directly verifiable information and Claude-based grader agents for aspects requiring interpretation, such as the level of certainty expressed in diagnostic conclusions. A grader is a separate AI-based evaluator that examines the agent’s answer against a predefined assessment rubric. The graders followed a predefined rubric but belonged to the same model family as the system being evaluated. Given the limited number of repetitions and the absence of independent human grading, the findings should be considered descriptive rather than evidence of statistically established performance differences.\n\nOne particularly interesting finding concerned how Claude selected its tools. When Predictive Maintenance MCP was available alongside Claude for Manufacturing, Claude autonomously discovered and used its diagnostic tools in every eligible run involving a vibration recording: 5 out of 5 in the pilot, 15 out of 15 in the main study, and 3 out of 3 in the Paderborn stage. In the single pilot case without a recording, PMM was not invoked. Importantly, the prompts did not explicitly instruct Claude to use PMM, and no custom integration code was required beyond configuring the MCP server.\n\nThe pump introduced at the beginning of this article is case P-105. Its synthetic maintenance history points toward a recurring outer-race problem, while the laboratory vibration recording suggests an inner-race defect. This makes it a useful example for examining how Claude handles conflicting sources of evidence.\n\nThe records make this a reasonable theory. A March work order found the outer race spalled, its surface flaking in the load zone, and the housing bore 0.04 mm oversize. The bearing was replaced, and the new housing that arrived in June was stored without being fitted. A worn bore lets the outer ring move, and a moving ring damages the outer race. There is one more source, represented by a vibration reading taken on the motor drive-end bearing at 10:10, twenty minutes before the stop. It is an MFPT recording whose sealed label is an inner-race fault.\n\nIn all three configurations of the repeated study, Claude followed the breakdown-diagnosis procedure. What differed was how it analyzed the vibration recording. Without PMM, Claude wrote its own Python code to perform envelope analysis, calculated the expected bearing fault frequencies from the bearing data sheet, and concluded that the measurement pointed to an inner-race defect.\n\nWith PMM connected, Claude instead used the server’s diagnostic tools. The bearing check identified a spectral line at 118.0 Hz, close to the expected inner-race frequency of 118.9 Hz, standing 40 dB above the local noise floor. Here, decibels describe how prominently a spectral line stands above its local background, not the overall loudness or vibration severity of the machine. A second line at 81.3 Hz, near the expected outer-race frequency, was 13.8 dB above the noise floor but lacked supporting harmonics and was not considered sufficient to confirm that fault pattern. The server therefore reported evidence consistent with an inner-race defect, while explicitly stating that the result did not constitute a confirmed fault.\n\nIn the examined PMM-assisted run, Claude then compared the technician’s hypothesis, the previous maintenance records, and the vibration analysis. It concluded that the earlier outer-race failure could not be confirmed as the cause of the current breakdown.\n\nThe subsequent recommendations were generated by the agent and were not independently validated as repair decisions. Claude suggested installing the replacement housing while the pump was already stopped, but cautioned against treating the housing as the confirmed root cause. It proposed inspecting the inner race and shaft seat first, followed by the outer ring and housing bore. It also identified an unresolved question that the vibration analysis could not answer: whether the March intervention concerned the motor’s drive-end bearing or the pump’s bearing, which would determine how relevant that maintenance history was to the current measurement.\n\nAcross five repetitions of each configuration, none of the 15 runs confirmed the technician’s original theory. All identified evidence consistent with an inner-race defect, matching the laboratory specimen’s fault label, although the degree of certainty expressed in the answers differed. Importantly, the technician’s hypothesis was challenged under both approaches to signal analysis, illustrating the contribution of the breakdown-diagnosis workflow itself.\n\nThe other two scenarios introduced different challenges. P-101 involved a running pump flagged during a routine vibration inspection. Although its recording was labelled as an inner-race fault, PMM identified spectral evidence supporting two possible fault locations: an inner-race component 42.3 dB above the local noise floor and an outer-race component 15.3 dB above it. Under the server’s decision rules, both patterns were supported, so neither was selected as the sole explanation.\n\nP-103 instead examined how Claude handled measurement history. The scenario described a pump whose vibration had progressively worsened over three months. However, the four recordings assigned to consecutive inspection dates actually came from separate laboratory measurements, not from repeated observations of the same bearing. This was explicitly documented in the case files. The experiment therefore tested whether Claude would preserve that important limitation when interpreting the apparent degradation trend.\n\nFigure 5 illustrates why this distinction matters, as measurements collected from different bearings cannot be treated as successive observations of a single component, even when they are assigned consecutive inspection dates.\n\nThe first comparison concerns how specifically Claude identified the fault location. This requires a distinction between recognizing evidence consistent with the laboratory label and selecting that label as the only possible explanation.\n\nUsing Python alone, Claude selected the single labelled fault type in 14 of 15 runs. In the remaining run, no signal analysis was completed because the test harness rejected a shell command. With PMM available, the corresponding count was 10 of 15, both with and without the additional evidence-reporting instructions.\n\nThe difference came entirely from P-101. In all five repetitions of each PMM-assisted configuration, the server found evidence supporting both inner-race and outer-race fault patterns. Claude therefore retained both candidates, including the laboratory-labelled inner-race fault, rather than selecting one.\n\nThe Python-only approach produced more specific conclusions, but a more specific conclusion is not necessarily a better-supported one. Conversely, retaining multiple candidates does not automatically make a diagnostic assessment more accurate. The comparison shows a difference in how the systems handled ambiguous evidence, not a demonstrated advantage in fault-detection performance. None of the configurations asserted a fault type inconsistent with the laboratory label.\n\nA more noticeable difference emerged in how Claude expressed the certainty of its diagnostic conclusions. When relying on its own Python analysis, Claude presented the identified fault type as an established fact in 12 of 15 runs. This happened in none of the 15 runs where PMM was used automatically, and in only one run when additional evidence-reporting instructions were provided.\n\nP-101 illustrates this distinction particularly well. In the four Python-only runs that successfully analyzed the recording, Claude identified an inner-race defect as the cause, generally treating the outer-race components as weaker evidence when they were mentioned. PMM, however, found that both inner-race and outer-race signatures satisfied its predefined diagnostic criteria. Its decision rules therefore prevented either candidate from being selected as the sole explanation, regardless of their relative spectral amplitudes.\n\nConsequently, Claude’s Python-based conclusions were more specific and matched the laboratory label, while the PMM-assisted responses retained both possibilities and expressed greater uncertainty. Neither approach establishes which interpretation would be preferable in a real maintenance investigation, where additional measurements and contextual evidence would be needed to resolve the ambiguity.\n\nI also examined the quality of the explanations produced by each configuration (Figure 7). AI-based graders identified 28 overclaims in the Python-only runs, compared with 17 when PMM was used automatically and 15 with the additional evidence instructions. Unsupported claims followed a similar pattern, with 20, 13 and 3 occurrences, respectively. However, factual contradictions showed a different trend: 5 with Python alone, 9 with PMM used automatically and 4 with the additional instructions. These are counts of individual claims across 15 runs per configuration, assessed by AI graders, rather than diagnostic error rates.\n\nFive of the nine factual contradictions observed when PMM was used automatically were related to the same issue: an operation that appeared to read data had actually modified the server’s local state. When a recording includes an asset identifier in its companion metadata, PMM’s load_signal tool also registers that measurement in a local asset ledger. This enables subsequent comparisons with historical measurements, but it means that loading a signal is not necessarily a read-only operation. None of the 15 runs in this configuration disclosed this update, and several responses explicitly stated that no data had been written or modified.\n\nThe integration also struggled to preserve information returned by the diagnostic tools (Figure 8). Claude explicitly reported PMM’s claim level in only 6 of 15 runs, although the AI graders considered its diagnostic wording appropriately cautious in 13. In other words, Claude often respected the uncertainty expressed by the server without communicating the corresponding evidence classification. Moreover, none of the 15 responses retained the numerical values of every spectral line evaluated by PMM.\n\nAn even more consequential issue emerged in P-103, where the synthetic maintenance history described a pump whose vibration had progressively worsened over three months. However, the four recordings used to construct that history came from separate laboratory acquisitions, not repeated measurements of the same bearing. This limitation was documented in the case files and was essential for interpreting the apparent trend.\n\nWithout PMM, Claude preserved this distinction in 4 of 5 runs. With PMM connected, whether used automatically or with additional evidence instructions, it was omitted in all five repetitions of each configuration. The resulting answers interpreted the sequence as evidence of degradation in a single bearing, even though the experimental data did not support that conclusion.\n\nAlthough PMM returned provenance information with the loaded recordings, the final answers did not preserve the distinction needed to interpret the synthetic history correctly. In this scenario, the two PMM-assisted configurations also produced more grader-identified overclaims than the Python-only configuration: 11 and 10, respectively, compared with 8.\n\nIn this context, provenance means knowing where a measurement came from, which component it represents and whether it can legitimately be compared with another recording.\n\nThe additional evidence-reporting instructions improved how Claude communicated the server’s results. With these instructions, all 15 responses explicitly included the diagnostic claim level, the numerical values of every tested spectral line, whether shaft speed was measured or declared, and the update to the local asset ledger. The number of unsupported claims identified by the graders also fell to just three across the 15 runs.\n\nHowever, these improvements did not resolve every issue. In P-103, none of the five responses preserved the warning that the apparent degradation history was assembled from separate laboratory recordings. In one run, Claude also placed a summary before the lockout safety reminder that the manufacturing skill requires at the beginning. The additional instructions increased the median response length by approximately 6–12%, depending on the case. Importantly, these instructions were added to the experimental prompt; they were not part of either the original Claude for Manufacturing skill or PMM.\n\nTo explore how the systems behaved on different vibration data, I conducted a final set of experiments using the Paderborn University bearing dataset [7]. Unlike the MFPT recordings, these included bearings that developed damage during accelerated lifetime testing rather than through artificially introduced defects. I selected three specimens, labelled as having inner-ring damage, outer-ring damage, and no damage. Each was analyzed once using Claude for Manufacturing with Python, PMM independently, and Claude for Manufacturing connected to PMM. These nine runs were exploratory observations, not an additional repeated comparison (Figure 9).\n\nTwo terms are important for interpreting these results. PMM assigns a strength grade to the evidence associated with a particular fault-frequency pattern, indicating how that pattern satisfies the server’s diagnostic criteria. Separately, its claim level constrains what can be concluded from the available measurements. Neither is a probability that a fault is present, and strength grades should not be used to rank competing fault locations. A strong spectral indication is still not the same as a confirmed mechanical failure.\n\nFor the specimen labelled with inner-ring damage, all three configurations identified evidence consistent with an inner-race defect. The outer-ring specimen produced a more ambiguous result. Claude using Python identified the outer race as the likely fault location, consistent with the laboratory label. PMM, however, detected three supported fault-frequency components, including an outer-race component 34.9 dB above the local noise floor and an inner-race component 20.9 dB above it. Interestingly, the server classified the outer-race evidence as *moderate* and the inner-race evidence as *high*. These strength grades are not probabilities or a ranking of competing diagnoses. Following the server’s decision rules, both PMM-assisted analyses therefore retained multiple possible fault locations rather than selecting one.\n\nThe specimen labelled healthy was particularly interesting. Its vibration recording contained strong impulsive components, and none of the three configurations concluded that the bearing was healthy. When used independently, PMM did not identify sufficient evidence for a specific bearing fault and explicitly noted that the absence of supported fault-frequency components does not establish that a bearing is healthy.\n\nClaude for Manufacturing using Python instead described the recording as containing a fault signature, while leaving a rolling-element defect and an electrical source as competing explanations. When Claude for Manufacturing used PMM, it acknowledged the server’s inconclusive result but still listed early bearing wear as a possible cause, at low confidence and tied with another explanation. The downtime record generated in that run went further, naming rolling-element damage while marking it as unconfirmed, despite the server not supporting that specific fault pattern.\n\nThis was only one execution, so it does not establish how frequently such behavior would occur. Nevertheless, it illustrates an important limitation of the combined workflow: **a diagnostic tool may appropriately withhold a conclusion, yet the agent can still introduce a more specific, insufficiently supported cause into the maintenance documentation.**\n\nConnecting PMM also affected the computational resources required to complete the investigations. Table 1 summarizes the aggregate token consumption and estimated costs across the 15 repeated runs of each configuration, with percentage changes calculated against Claude for Manufacturing using Python alone. Figure 10 provides the corresponding per-case distributions of execution time, input tokens and estimated costs.\n\nThe configuration with additional evidence-reporting instructions also recorded lower input-token consumption and estimated cost than PMM used automatically: approximately 18% fewer input tokens and 7% lower estimated cost. These differences were observed across separate runs, however, and cannot be attributed to the additional instructions alone.\n\nThe token figures may initially seem counterintuitive. Without PMM, Claude generated Python code to perform vibration analysis. With PMM, it could invoke existing diagnostic functions instead. Yet the Python-only configuration recorded fewer input tokens.\n\nThe distinction lies partly in how model interactions are counted. Code written by Claude initially contributes to output tokens, although it can also become part of the input context in subsequent requests. When PMM is connected, tool discovery, selected tool definitions, returned results and the evolving investigation can contribute additional input context. The reported totals also include cache creation and repeated cache reads, not just newly processed text. Consequently, a 73% increase in recorded input tokens does not imply 73% more unique information, nor an equivalent increase in estimated cost. In this experiment, the estimated cost increased by 23%.\n\nThe experiments did not isolate the contribution of each mechanism, so the additional input consumption cannot be attributed to a single source. More importantly, these figures measure the resources consumed during the investigations, not the effort required to develop or validate the underlying diagnostic methods. Claude can generate an analysis procedure for each investigation, but the resulting code and analytical choices are not guaranteed to be identical across independent runs. PMM instead executes predefined implementations and diagnostic rules that can be tested and reviewed separately from the language model. Whether this consistency provides enough operational value to justify the additional model cost requires further investigation.\n\nThe differences in execution time were less straightforward. Although the PMM configurations had lower total elapsed times across the repeated runs, their per-case ranges overlapped with the Python-only configuration in most scenarios. A clear separation was observed only for P-105, where both PMM configurations completed the investigations faster. Similarly, estimated costs showed non-overlapping ranges only for P-103, where both PMM configurations were more expensive, while input token usage showed a clear separation only for P-103 with PMM used automatically.\n\nThese observations illustrate the operational trade-offs of introducing specialized tools into an agent’s workflow. In my experiments, PMM increased the amount of input context and the estimated API cost, while reducing the amount of generated output. However, it did not consistently increase execution time. These results are specific to the three scenarios, the Claude model, and the client version tested, and should not be generalized to other workloads without further evaluation.\n\nThe experiments also highlighted several issues that are relevant beyond this particular combination of tools.\n\nThe value of a specialized diagnostic toolkit cannot be assessed solely by comparing whether a language model can reproduce an individual calculation. Claude showed that it could write Python code for envelope analysis and calculate bearing characteristic frequencies. PMM offers a different way to organize the same analytical work.\n\nWith PMM, the numerical methods and diagnostic decision rules reside in a dedicated software component rather than being generated as part of each investigation. An engineer can inspect how a fault-frequency match is evaluated, which thresholds are applied, why evidence is considered sufficient or inconclusive, and how those rules change between software versions. The agent remains responsible for selecting the appropriate tools and interpreting the results, but the underlying implementations can be reviewed and tested independently.\n\nThis separation becomes relevant when the same diagnostic procedure must be applied across different machines, recordings or investigations. Claude can generate suitable code, but independent sessions are not guaranteed to produce identical implementations or analytical choices. PMM instead provides reusable procedures whose behavior can be examined under controlled conditions. That is an architectural advantage, not proof that its diagnostic decisions are always correct.\n\nThe repeated experiments provide one concrete illustration. On P-101, PMM consistently retained multiple supported fault hypotheses, while Claude using Python generally selected the laboratory-labelled inner-race fault as the cause. PMM-assisted responses also contained fewer grader-identified overclaims. The benefit observed was therefore not greater fault-label agreement, but a more constrained interpretation of the available diagnostic evidence.\n\nPMM’s broader capabilities extend beyond this particular comparison. Its implemented tools cover signal analysis, severity assessment, measurement history, anomaly detection, prognostics, documentation retrieval and reporting. These functions could support investigations that extend well beyond a single bearing spectrum, although their combined operational value was not evaluated here.\n\nThe additional estimated model cost was approximately $0.20 per investigation in the repeated comparison, with considerable dependence on the scenario. This figure captures neither the cost of validating dynamically generated analytical code nor the potential value of reusable diagnostic methods in an industrial workflow. Those factors remain open questions, alongside opportunities to reduce tool-context overhead and improve how diagnostic evidence reaches the final report.\n\nThe first concerns how diagnostic evidence moves between components. MCP made PMM’s tools available to Claude, but the manufacturing skill and the server did not define a shared procedure for transferring diagnostic evidence into the final maintenance report. For example, the skill’s downtime record includes a field for the most likely cause, but no corresponding structure for the measurements supporting that conclusion. Meanwhile, PMM provides diagnostic outcomes, claim levels and numerical evidence without controlling how another workflow presents them.\n\nAs a result, Claude had to decide which information to retain in each response. The additional instructions substantially improved this behavior, suggesting that an explicit *evidence contract* between the skill and the server could be useful. Such a contract would specify which diagnostic findings must be preserved, how uncertainty should be expressed, and what information about measurement conditions and provenance must accompany each conclusion. However, the unresolved P-103 case shows that simply adding reporting instructions is not enough to guarantee that the original context survives.\n\nA second issue concerns operations that modify data. Loading a recording through PMM could update its local asset ledger, even though the operation appeared to be primarily about reading and analyzing a signal. Claude did not disclose these updates until the experimental instructions explicitly required it. This suggests that tools should communicate their side effects clearly and that the surrounding workflow should preserve that information in its final output, rather than relying entirely on the model to recognize and report it.\n\nThe third issue is how instructions from different components interact. PMM’s server instructions tell Claude not to recommend actions beyond those supported by tool outputs, while the manufacturing skill requires a proposed repair path for each candidate cause. These instructions can create conflicting expectations when the systems are combined. In a separate probe, I also observed that Claude passed only approximately 4,000 characters of PMM’s 13,056-character server instructions to the model. This was a single observation, but it reinforces the importance of keeping server instructions focused on the tools and evidence they provide, while making responsibilities between components explicit.\n\nThe fourth observation concerns experimental repeatability. Some behaviors that appeared in the initial pilot did not persist in the repeated experiments. For example, the pilot suggested that the integrated configuration omitted a change assessment and introduced an incorrect safety flag. Neither observation was consistently reproduced across five repetitions. This illustrates why conclusions about agent behavior should not rely on isolated runs, particularly when evaluating workflows involving multiple tools and decisions.\n\nReturning to the comparison, both approaches were capable of analyzing the vibration recordings. Claude using Python identified the laboratory fault label at least as often as the configurations using PMM. The main contribution observed from the specialist server was different: it supplied a predefined diagnostic procedure, explicit criteria for interpreting spectral evidence, and structured assessments that could be examined independently of Claude’s reasoning.\n\nP-101 provides a useful example. Whenever the server evaluated that recording under the tested conditions, it returned the same outcome: evidence compatible with multiple fault locations, without selecting one as the confirmed explanation. This consistency is valuable when diagnostic results need to be reviewed, compared or documented, even if it does not necessarily produce a more specific conclusion. Whether PMM’s other capabilities provide additional benefits to a complete maintenance investigation remains to be evaluated.\n\nThese findings must be interpreted within the limitations of the study. The vibration recordings came from laboratory datasets rather than operating industrial machines, and the surrounding maintenance information was synthetic. The Paderborn recordings were obtained through archived copies of the official downloads. When two initially selected recordings could not be retrieved completely, they were replaced with readable recordings from the same respective bearings and operating conditions. The recordings ultimately analyzed were complete.\n\nThe synthetic context also influenced some results. In P-103, all 15 repeated runs raised a safety flag because the supplied maintenance records included an unresolved electrical fault on a different pump. This reflects an ambiguity in the constructed scenario rather than evidence of a fault on the machine being investigated.\n\nThe evaluation included only five repetitions per configuration and scenario, used AI graders without independent human assessment, and covered a single model and Claude Code version. Several manufacturing skill capabilities, including photograph analysis, manufacturer troubleshooting manuals, live connectors and batch investigations, were not tested. Most PMM capabilities were likewise outside the scope of the experiments. I also identified implementation issues in the server, including a vibration severity result that incorrectly labelled a horizontal measurement as vertical. None of these experiments establishes diagnostic accuracy, reduced repair times, or economic benefits in an operating plant.\n\nThe experiments point to six practical conclusions:\n\nReturning to the pump that opened this investigation, both approaches were able to challenge a plausible diagnosis based on maintenance history. Claude demonstrated that it could perform meaningful vibration analysis by generating Python code. PMM demonstrated a different capability, that is providing predefined diagnostic methods and explicit evidence rules that the agent could invoke without recreating them.\n\nFor an engineer developing a repeatable maintenance workflow, that distinction matters. The objective is not simply to obtain a plausible answer once, but to understand how it was produced, examine the criteria behind it and apply comparable procedures across future investigations. PMM provides a software foundation for this separation of responsibilities, while Claude connects the technical findings with equipment history and maintenance decisions.\n\nThe experiments also showed that deterministic diagnostic tools do not automatically make the complete agent reliable. Evidence can still be omitted, misunderstood or presented with greater certainty than the measurements justify. Making these systems dependable will require better evidence transfer, clearer tool responsibilities and validation with real industrial data.\n\nThe value of specialist tools is therefore not that language models cannot perform the calculations themselves. It is that industrial users may need diagnostic methods they can inspect, test and reuse independently of the agent generating the final explanation.\n\nThat is where the next stage of development and evaluation should focus, using field recordings, maintenance histories and independent engineering assessment.\n\n**Data.** MFPT bearing data set (E. Bechhoefer), via The MathWorks, Inc., CC BY-NC-SA 4.0. Paderborn University KAt Bearing DataCenter (C. Lessmeier et al.), CC BY-NC 4.0. No recording is reproduced in this article.\n\n**Software.** Claude for Manufacturing plugin, Anthropic, Apache License 2.0. Predictive Maintenance MCP, MIT License. Product names are used only to identify what was tested.\n\n[Testing Claude for Manufacturing with Predictive Maintenance MCP](https://pub.towardsai.net/testing-claude-for-manufacturing-with-predictive-maintenance-mcp-1ca86d0c86b4) was originally published in [Towards AI](https://pub.towardsai.net) on Medium, where people are continuing the conversation by highlighting and responding to this story.", "url": "https://wpnews.pro/news/testing-claude-for-manufacturing-with-predictive-maintenance-mcp", "canonical_source": "https://pub.towardsai.net/testing-claude-for-manufacturing-with-predictive-maintenance-mcp-1ca86d0c86b4?source=rss----98111c9905da---4", "published_at": "2026-11-22 14:01:02+00:00", "updated_at": "2026-11-22 14:17:01.348380+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "agent-protocols", "ai-products", "machine-learning"], "entities": ["Anthropic", "Claude for Manufacturing", "Predictive Maintenance MCP", "LGDiMaggio", "Model Context Protocol", "Microsoft 365", "Oracle NetSuite", "Infor M3"], "also_reported_by": [], "takeaway": "Supplying diagnostic calculations and evidence rules through a dedicated MCP toolkit, rather than letting the agent generate signal-processing code per investigation, is what the study compares on diagnostic accuracy, uncertainty, evidence preservation and computational overhead.", "recommendation": {"advice": "Route industrial diagnostic work through dedicated MCP tools with fixed standard procedures instead of letting the agent write its own analysis code, so conclusions stay reproducible and independently verifiable.", "playbook": "agent-architecture", "playbook_url": "https://wpnews.pro/playbooks/agent-architecture"}, "alternates": {"html": "https://wpnews.pro/news/testing-claude-for-manufacturing-with-predictive-maintenance-mcp", "markdown": "https://wpnews.pro/news/testing-claude-for-manufacturing-with-predictive-maintenance-mcp.md", "text": "https://wpnews.pro/news/testing-claude-for-manufacturing-with-predictive-maintenance-mcp.txt", "jsonld": "https://wpnews.pro/news/testing-claude-for-manufacturing-with-predictive-maintenance-mcp.jsonld"}}