On November 12, 2026, Anthropic released "Claude for Manufacturing," a plugin that supports manufacturing tasks such as quoting, quality management, and maintenance. It bundles 17 skills and lists integrations with Autodesk Fusion, Infor, Microsoft 365, and Oracle NetSuite. The idea is to use existing business systems and documents to produce quotes and candidate root causes for staff to review.
This is not an announcement of a new manufacturing-specific model. It is an effort to combine Claude with defined work procedures and connection points. What matters when adopting it is not just the names of the systems it connects to, but where the necessary records come from, how far the process runs automatically, and who executes changes.
Tasks that run on CSV versus tasks that need a connection
The 17 skills break down into 12 standalone skills, each handling an individual job, and 5 workflows that combine them. They cover sales, quality, and production operations, and also support procurement, design, and initial setup. For quoting, for example, Claude references the customer's request and the costs of similar past jobs; for quality management, it looks up past nonconformances. For maintenance, it uses sensor exports and repair notes. It also refers to photos and manuals, ranking candidate causes of a failure to assist the technician's judgment.
According to the public implementation's documentation, 11 of the 12 standalone skills are designed to work from exports such as CSV files, without a direct system connection. It also describes a way to periodically export data from systems that lack a dedicated connector and place it in Microsoft 365. File-based input is not just an emergency fallback; it is an intended input path.
The exception is "design-space-explore," which generates design candidates. It requires a live CAD connection, such as the Fusion desktop app. By contrast, a report comparing design candidates can be built from a hand-filled list of candidates. Generating designs and comparing existing ones require different connections.
Fusion is the only integration whose connection settings are written inside the plugin; Microsoft 365, NetSuite, and Infor M3 are configured on the Claude side. To use Fusion, you need to enable the MCP server in the local desktop app and have a document open. The four named integrations do not all work with the same setup effort.
When combining multiple records, the design relies on existing IDs for parts, manufacturing orders, and sales orders, as well as lot, equipment, and supplier IDs. Records that fail to match are reported, missing values are not treated as zero, and stale exports are flagged prominently. These mechanisms help users judge how fresh or incomplete the source material is; they are not a measured guarantee of output accuracy.
Check what data can be read and what can be changed, not just what connects
The table below, based on official materials, compares CSV input, Infor M3 reads, a separate Infor plugin, and equipment connection via MHS by input, execution, and approval conditions. It shows that what can be read and changed differs by connection path.
| Path | Input and execution conditions | Change and confirmation boundary |
|---|---|---|
| CSV and other exports | 11 of the 12 standalone skills are designed to run without a direct connection | Check for stale data, missing values, and unmatched records |
| Infor M3 integration in Manufacturing | Reads from M3, but routings and bills of materials must be exported | Changes to M3 are made by a person; the plugin drafts the request |
| Separate Infor plugin | Used according to Infor's features and permissions | Can also handle purchase orders and production schedule changes under human approval and permissions |
| Model Hardware Standard | A separate research preview that handles programmable equipment via drivers | Describes safety limits; expert supervision is required |
The Infor product names deserve particular care. The ERP options in Manufacturing are NetSuite or Infor M3; Infor LN and CloudSuite Industrial are separate products treated as custom integrations. Even if you can connect to M3, not all the information a quote needs will arrive.
Standard costs, which the M3 integration does not retrieve, are first filled in with the most recent purchase price, then with exported or supplied cost figures. Past quotes must be prepared from documents, and routings and bills of materials, which the M3 server does not expose, must also be exported. What remains is the work of separating, task by task, what can be fetched automatically from what must be gathered separately.
For example, if you use the costs of similar jobs in a quote, you should not finish the work by looking only at M3 records; you need to confirm that past quotes and routings are also in hand. If bills of materials and routings are supplied as files, the point in time when they were exported also becomes a factor in judgment. These conditions are easy to miss when comparing feature counts, but they are essential for staff to review the output. Whether the inputs can be assembled also depends on how well each company maintains its records.
The meaning of numbers must be aligned too. According to the implementation documents, M3 search results are cut off at about 30 rows, sometimes without notice. The design therefore re-fetches with narrower conditions when results reach 30 rows or more. Time units are also checked for each work center, and values are not converted to cost when the unit is unclear. This illustrates that a successful connection alone does not tell you whether the input is sufficient.
The M3 integration on the Manufacturing side is limited to specified reads and does not modify M3 even with approval. It drafts a change request, and a person executes it in M3. The separate Infor plugin, by contrast, can also handle purchase orders and production schedule changes under human approval and Infor-side permissions. Treating the two as one feature risks misjudging who does what after adoption.
Reports run automatically; corrective-action completion is confirmed on the floor
Where human review is needed also varies by the type of deliverable. The skill for investigating quality problems does not merely list candidate causes; it presents evidence for and against each candidate, along with tests that help tell them apart. In 8D reports, which organize root-cause investigation and recurrence prevention into eight steps, D6 (implementation and verification) and D8 (closure) are designed to be left blank. Producing a document is a different stage from having a corrective action carried out and verified.
Meanwhile, the daily production report workflow runs without an approval step. It does not rewrite the system of record that manages the authoritative business data, and instead updates a dashboard in the location where deliverables are stored. Automatic startup requires a separate schedule, such as in Cowork. A report that compiles on a schedule and work that changes official records operate under different boundaries.
Anthropic says the plugin is especially well suited to make-to-order manufacturing that handles individual quotes, drawings, and routings. It says daily production reports, failure diagnosis, and procurement can also be used in process manufacturing such as food, consumer goods, chemicals, and steel. However, it states that the quoting and quality procedures need to be adapted to batch records and each company's operations.
The public repository states that the code is a reference implementation, is not subject to active maintenance, monitoring, or acceptance of outside contributions, and comes with no warranty. This notice concerns the terms of the public code. The announcement points to adoption on Team and Enterprise plans, and the FAQ describes paid plans, including Enterprise via AWS Marketplace. Whether there are additional fees, Japanese-language support, and any terms specific to Japan are not stated in the official announcement or elsewhere.
Keep this separate from research on operating equipment
Claude for Manufacturing itself does not operate equipment. Separately, on August 27, Anthropic announced a research preview of the "Model Hardware Standard" (MHS). It is a mechanism in which drivers describe measurements, settings, and safety limits, handling equipment that can be operated programmatically through MCP and similar means. Equipment without a programmatic interface is outside its current scope, and limits in spatial and physical reasoning mean expert supervision is required.
Taken together, the practical unit of adoption looks to be a single task rather than the whole factory. First confirm the range and freshness of the records that can be obtained and the units of the numbers, then decide whether reports will update automatically or whether change proposals will be handed to a person. The public materials show feature design and usage examples; they do not present measured accuracy, labor savings, or return on investment at independent factories. Whether those conditions can be verified at each site will shape the adoption decision.