Claude for Manufacturing’s onboarding Skill sets up a plant profile by asking about the systems, files, terminology and operating context the other Skills need. Start with “set us up,” identify the sources you can actually provide, and review the resulting configuration and gaps. You can defer unknown answers, but downstream work may be limited until they are resolved.
Anthropic’s onboarding Skill is part of a reference implementation. The instructions describe the intended workflow; they do not prove that it has been validated in your plant. The plugin README describes installation and the four named connectors.
Install the plugin and begin setup
Anthropic’s release explains how to get started on Team and Enterprise plans; its product page describes availability across paid plans. Users can add available plugins themselves where their organization permits it. Administrators may restrict availability or manage deployment, so check your organization’s settings and permissions. You do not need a live ERP or historian connection for an initial export-based evaluation.
In Claude, open Customize → Plugins → Discover, select Claude for Manufacturing, and choose Add. In Cowork, open the Cowork tab before Customize. These are Anthropic’s documented plugin steps; the labels and access available to your account may differ. See Anthropic’s plugin instructions and manufacturing availability. This guide has not been run against a live plant.
Start a session and enter:
Set us up for one plant. We use a maintenance system and receive daily
production, downtime and historian exports. Please identify what you can
access, ask about our shift boundaries and source clocks, and list what is
missing and which Skills each gap affects. Do not invent a connection.
Use your actual system names and location when you run it. The setup interview can use supplied documents and ask you to confirm their contents. Provide credentials through the service’s approved connection flow, not in the conversation. If an answer is unknown, say so and assign it to an owner; don’t substitute a guessed time zone or asset mapping.
A first-run review example
For a fictional plant with a 06:00 production-day boundary, a production CSV and a maintenance export, a useful setup review would identify each source separately, name its clock and time zone, describe the shift boundary, and flag a missing historian export. If a shared event cannot be used to check clock skew, mark alignment unverified. This is a review checklist, not a transcript or a promised output format. Ask the production and maintenance owners to correct the profile before using it for cross-source analysis.
A named asset matters as much as a named source. Industrial data contextualization explains why an asset, signal and unit need to be associated before a measurement can be interpreted.
Prepare the first session
- List systems by role. Record the ERP, CAD, maintenance and document systems, plus the MES, historian, quality and planning sources you expect to use. Include the owner and export contact for each.
- Bring representative material. A sample production export, maintenance work order, current drawing or procedure and existing report template help the Skill identify what is already known. Use approved, appropriately scoped records.
- Describe time and shifts. Identify each source’s time zone, whether timestamps include a zone, the shift boundary and production-day start. A source clock is a separate question from the site’s local time.
- Review the configuration. Confirm the plant’s names for assets, lots and work orders; verify document interpretations and note unknowns. Review the gaps and the affected Skills before relying on an output.
The documented connection paths are Microsoft 365, Oracle NetSuite or Infor M3, and Autodesk Fusion. Fusion is declared by the plugin and requires its desktop application and local MCP server. The other connections are configured in Claude. Setup tests reachability; listing a system does not create its credentials or establish permission. Other systems can be named and served by exports while their connector options are assessed.
Check clocks before correlating events
The Skill records clocks per source and performs a skew self-check using a shared event where available. If there is no shared event to test, configuration remains declared but unverified. An unverified clock should appear in the gaps list; a later investigation should not treat the declared time zone as proof that two clocks agree.
For example, a maintenance work order stamped 02:10 and a historian event stamped 02:08 may appear related. First establish whether those timestamps represent the same time zone, whether each system’s clock was checked, and whether the work-order timestamp describes the failure or a later entry. Onboarding captures context for that decision; it does not make an uncertain event time precise.
What the result should tell you
Expect a plant context and configuration profile, with missing information tied to the Skills it limits. A change in shift boundary, a new site or a replacement source warrants a configuration update. Multi-site context matters because one plant’s production day must not silently become another’s.
For an OT team evaluating direct access to time-series data, TDengine’s TSDB MCP documentation describes read-only query tools and a Claude Code configuration example. That is a separate connection to evaluate: Anthropic’s onboarding instructions do not establish a tested TDengine integration or automatically map its results into the manufacturing dataset.
Before a real report or investigation, try one bounded example and inspect the source coverage, clock status and gap list with the source owners. Do not treat a completed onboarding conversation as evidence that the records are current, complete or correctly joined.


