Claude for Manufacturing is Anthropic's plugin for plant work: 17 skills that draft quotes, NCRs, 8Ds, breakdown diagnoses, supplier requests and the morning production report, with a person approving each one. It runs on nightly exports from your ERP, MES, quality and maintenance systems, and its own rules assume those exports arrive on time and join on the same part numbers. Installing it takes minutes. Most of the setup is data work.
Anthropic released Claude for Manufacturing on 12 November 2026. The plugin's source is public on GitHub, so we read all of it: every skill, every reference file and the shared rules underneath.
We read it with one question in mind, because it's the question our team answers for a living: what does this need from a company's data before its first draft is worth reviewing? The answer is more specific than the launch coverage suggests. The skills are thoughtful, and unusually honest about what they need from you. This guide covers what's in the plugin, what we noticed reading the source, how to set it up, and the prompts to start with.
What the plugin contains
A plugin is a bundle of skills and connectors you install into Claude. Skills are written instructions that tell Claude how to do one kind of work, step by step. Connectors let it read your systems. Once it's installed, nobody has to remember skill names. You describe the job in plain language, “an RFQ came in” or “cell 3 is down”, and the right skill picks it up.
Claude for Manufacturing has:
- 17 skills. Twelve do one job each. Five are chains that run several of those in order and stop for a person at each decision.
- Four connectors. Microsoft 365 for documents, mail and exports. Oracle NetSuite or Infor M3 for the ERP. Autodesk Fusion for CAD and CAM.
- One rule underneath all of it. It drafts, and your people decide. Quotes, NCRs, 8Ds and supplier requests all end with a person's review, and nothing is sent, applied or written without a yes.
It's available on paid Claude plans. Anthropic's announcement names Team and Enterprise. An admin installs it from the plugin directory, and then anyone with access says “set us up”.
Infor also released its own plugin alongside it. Anthropic's plugin only reads from Infor. Infor's can make changes in Infor once you approve them, like releasing a purchase order to another supplier when a steel shipment slips.
The 17 skills and what to say to each
The twelve single-job skills, grouped the way the plugin groups them:
| Skill | Say | What you get |
|---|---|---|
| Onboarding | “Set us up” | Tests your connectors, reads your documents, learns your exports, shifts and vocabulary, and hands back a list of what's missing |
| Data consolidate | “Pull the data together” | One workbook joined across your systems and exports, plus a report of every record that didn't match |
| Custom connector | “We use SAP” | The best available way to reach a system that has no connector, or an export specification for your IT |
| Customer inquiry intake | “An RFQ came in” | The RFQ in one form, every field sourced, what's missing listed, and a bid or no-bid recommendation |
| Quote build | “Price this job” | A quote that shows its math: material, labor, machine time, setup per quantity break, margin on its own line |
| NCR write-up | “We found bad parts” | A nonconformance cited to the requirement it violates, how many parts are suspect and where, and whether it has happened before |
| Root cause and CAPA | “Write the 8D” | Ranked candidate causes with evidence for and against, the cheapest test to tell them apart, and a draft 8D |
| Supplier quote request | “Source this part” | The RFQ package for suppliers, one draft email each, bids read into one shape and compared per quantity break |
| Breakdown diagnosis | “Cell 3 is down” | Likely causes ranked from sensor data, repair notes, photos and the manual, safety first, and the cheapest check to run |
| Daily production report | “Morning report” | Output against plan, OEE split into its three parts, downtime and scrap in money, and today's exceptions |
| Design space explore | “Generate variants” | A batch of variants of your existing Fusion model, after you approve the count |
| Trade study report | “Which design should we pick?” | A ranked, cited comparison with a sensitivity analysis and a decision memo |
And the five chains, which carry a job across several of those skills:
| Chain | Runs |
|---|---|
| RFQ to quote | Inquiry intake, data consolidate, a design step in Fusion, then quote build. Four points where a person approves |
| Inspection to CAPA | NCR write-up, then root cause and CAPA |
| Morning plant pulse | Data consolidate, then the daily production report, built to run on a schedule |
| Breakdown to next report | Breakdown diagnosis, then the next production report |
| Supplier escape to alternate source | NCR write-up, root cause and CAPA, then a supplier quote request to find another source |
Notice how many of these start with data consolidate, or quietly depend on it. That skill is where the plugin meets your systems, and most of the rest of this guide is about what it finds there.
Seven things in the source
1. It's built to run on a folder of CSVs
This is the most important line in the whole plugin: “Scheduled exports are the designed path, not a fallback.” Eleven of the twelve single-job skills work with no connector at all, from a folder of nightly exports. The plugin's notes put it plainly: a plant that can't give you an API “can almost always give you a nightly CSV, because it is already producing one for somebody's spreadsheet.”
The exception is design space explore, which needs Autodesk Fusion open on the desktop. There is no path for SolidWorks, Creo, NX or Inventor, because none of them publishes an MCP server, the protocol Claude uses to talk to other software.
For whoever owns the data, this cuts both ways. Nobody has to build an API before the first run. But a folder of exports is only as reliable as the jobs that fill it, and the plugin now depends on every one of them.
2. Its most important rule is about missing data
One shared file every skill follows is called “Absent is not zero, and stale is not current.” It calls this “the single most common way a manufacturing skill produces a confident wrong answer”:
“A file did not land, a row did not match, a field came back empty, or an export stopped running three weeks ago. The skill reads that hole as a fact about the plant rather than a fact about the data.”
So every report states up front how fresh each source is. If last night's export didn't land, that goes at the top, and only a safety issue goes above it. Anyone who has run a plant on exports will recognize the failure. The export didn't break. It just stopped, quietly, and the dashboard kept going.
We designed Databasin's ingestion around the same failure. A change column that hasn't moved in 90 days while its neighboring columns keep updating is flagged as stale instead of trusted, and a file pattern that matches no files won't save. The plugin will tell you when a source went quiet. Catching it before the morning report is the job of whatever produces the file.
3. It expects your systems to disagree with each other
The plugin assumes a plant has “five to eight systems bought at different times by different departments, and the same physical thing carries a different identifier in each one.” Its own example is one part:
| Where | How the part appears |
|---|---|
| ERP item master | PN-4471-B |
| MES production export | 4471B |
| Quality inspection export | 4471-B Rev 2 |
| Maintenance spares list | Bracket, Upper, 4471 |
Same part, four names, and none of them match as they are. Everything the plugin builds hangs off six identifiers: part number, production work order, sales order, lot, asset and supplier. It even warns that a maintenance “work order” and a production work order are different records, and joining one to the other “produces nonsense.”
A data team would call the fix a crosswalk: a maintained table that maps each system's identifier to one shared key, with an owner who adds new parts as they appear. Few mid-sized plants have one. Until someone builds it, data consolidate will list the same unmatched rows every time it runs.
4. Two ERPs connect out of the box. Yours may not be one of them
The ERP connectors are Oracle NetSuite and Infor M3. Every other system goes through the custom connector skill, which works down a ladder from a vendor's own connector to a scheduled export. Its notes on what plants actually run are blunt:
- SAP ECC: “Expect a project, not an afternoon.”
- Global Shop, JobBOSS and E2: “Export is often the pragmatic answer.”
- And the line every mid-sized manufacturer will recognize: “A large share of what's running in mid-market plants is on-prem, a decade old, and has a SQL database but no API at all.”
Even on Infor M3, the plugin gets no standard cost, quote history or routings, so the quoting skills fall back on exports for those.
That last quote describes a lot of the connection work we do. When a system has a SQL database and no API, the dependable route is usually a read-only connection to the database itself, scheduled and monitored, rather than a report someone set up in the application years ago and nobody has checked since.
5. Its quotes show their math
Quote build prices material at what you actually paid on a named purchase order, not a standard cost. It spreads setup across every quantity break so the price curve explains itself, and puts margin on its own line. On the internal sheet, every operation says where its hours came from: a current routing, actual hours from closed jobs, the design, or first principles. It flags actuals that look like standards that were backflushed into the job rather than measured. None of that working reaches the customer's copy.
Each of those sources is a dataset: purchase order history, closed-job labor, routings. A quote is only as good as the oldest of the three.
6. It won't finish what hasn't happened yet
The 8D skill drafts D1 to D5 and D7, and leaves D6 and D8 empty, “because those need something to actually have happened.” The supplier skill compares bids and recommends one, and never awards. A quote goes to your estimator, never straight to the customer.
The shared file on approvals is candid about what testing found: “That rule held under test: no write reached any system. What gave way was the edge of the yes.” A yes to one fix got stretched over four fixes nobody had seen. So the rule now is that a yes covers exactly the item that was shown, and “carry on” or “set us up” moves the work forward without approving anything.
7. The morning report is the one thing that runs on its own
The morning plant pulse is the only chain with no approval points anywhere in it, because it writes to no system of record. Its dashboard lives privately in your own Claude account. The plugin says a supervisor otherwise spends 60 to 90 minutes a day putting this report together by hand.
A plugin can't wake itself up, though. You schedule it yourself, as a scheduled task in Claude Cowork or Claude Code. The plugin is honest about this too: whether a scheduled run can update the dashboard on its own hasn't been verified yet, so check the first few mornings.
Setting it up, step by step
The plugin's onboarding is good, and it never blocks: “I'd have to ask” is accepted as an answer, and anything you skip goes on a list. But it can only work with what you bring it. Most of the setup happens before you say “set us up”.
Step 1. Pick one process you already measure
Anthropic's own advice is to start with one process you measure, and we agree. Pick the one where you can already say how long it takes and how often it goes wrong:
| If your plant | Start with |
|---|---|
| Quotes custom jobs from drawings | RFQ to quote |
| Owes customers NCRs and 8Ds | Inspection to CAPA |
| Builds the morning report by hand | Morning plant pulse |
| Makes food, chemicals or steel | The morning report, breakdown diagnosis and the supplier skills, which Anthropic says fit process manufacturers today |
Write down how long that process takes today, and how often it goes wrong. You'll need the number in step 6.
Step 2. Gather the documents before you start
Onboarding asks for a pile of documents up front, reads them, and turns the rest of the interview into “confirm or correct”. The more you bring, the fewer questions you answer. For the process you picked, gather what applies:
- Production and maintenance: an MES or production export, the maintenance history from your CMMS, your machine or asset register, and a sensor export if you have one.
- Quality: a recent NCR, a recent 8D, inspection records, and your quality manual or ISO, AS9100 or IATF certificate.
- Quoting: a recent quote you sent, your supplier RFQ package, and your standard supplier quality clauses.
- Plant setup: your shift pattern, a part master and routing export, setup times or cutting data, and the approved supplier list. The approved supplier list is always asked for, because the ERP knows your vendors but not which are approved for what.
Step 3. Set up the export folder, and give every export an owner
For every system without a connector, set up a nightly export into one shared folder in Microsoft 365. Give each file the same name every night, with the date in it, and write down which job produces it and who is told when it fails. The plugin will flag a stale file, but someone still has to fix it, and that goes faster when the alert reaches them before the first shift does.
Where a system has a database you can reach, export from the database with a scheduled query rather than from a screen in the application. The fields stay the same from night to night, and a failed run is a failed job you can see, instead of a file that simply doesn't appear.
If you don't know how to get data out of a system, ask the plugin. The custom connector skill checks whether the vendor already publishes a connector before suggesting anything harder:
We run [SYSTEM] for [WHAT IT HOLDS, E.G. MAINTENANCE TICKETS]. It isn't one of your connectors. Work out the best way for you to read it, starting with whether the vendor already publishes an MCP server. If the honest answer is a scheduled export, write the export specification for our IT: which fields, how often, which folder, the file name, and how we'll know when it stops running.
Step 4. Agree on your plant's words, numbers and keys
Onboarding asks what your plant calls things. A job, a work order or a traveler? A cell, a line or a work center? It also asks for your labor and machine rates. Agree on these before the conversation, not during it, especially the rates, which end up in every quote.
Then run data consolidate once on your exports, before you rely on anything built on top of them. The unmatched-records report is the most useful thing you'll read in your first week. Treat it as the specification for your crosswalk:
Pull the data together from our export folder for [DATE RANGE]. Join on part number, production work order, sales order, lot, asset and supplier. Then report every record that didn't match, why it didn't, and what you'd need from us to fix it. Tell me how fresh each source is. Don't apply any corrections without showing me each one first.
Step 5. Install it and say “set us up”
An admin installs the plugin from Claude's plugin directory. Turn on the Microsoft 365 connector, and NetSuite or Infor M3 if you run one, in Claude's connector settings. For Fusion, open the desktop app and go to Preferences, General, API, and turn on the Fusion MCP Server. Then start the conversation:
Set us up. We make [WHAT YOU MAKE] at [NUMBER] plant(s). The first process we want to run on this is [PROCESS]. I'm attaching the documents we have: [LIST]. Our scheduled exports land in [FOLDER]. At the end, give me the gaps list: what's still missing, which skills each gap limits, and the options for connecting every system you can't reach yet.
Read the gaps list carefully. It's your to-do list for the next two weeks, and most of it will be data work.
Step 6. Run it next to the old way
For the first two weeks, run the chain alongside the way you do it today and compare the two. Use the number from step 1. Where the draft is wrong, find out whether it was the skill or the data. Usually it's the data, and the fix belongs in the export or the crosswalk, not the prompt.
Prompts to start each chain with:
An RFQ came in, attached. Read it, fill in our intake form, and list what's missing from the drawing. Tell me whether you'd bid or not, and why. If we bid, build the quote and show the math.
We found bad parts. [WHAT WAS FOUND, WHERE, THE LOT, THE INSPECTION RECORD.] Write up the nonconformance against the requirement it violates. Check whether this has happened before and whether the last fix held. Then draft the 8D.
Step 7. Schedule the morning report
Once the morning report is right by hand, schedule it as a task in Claude Cowork or Claude Code, to run before the first shift. Schedule it after your nightly exports finish, with enough margin for a slow night:
Run the morning production report for yesterday.
Saying “morning report” yourself at 6:15 produces exactly the same page. That's a fine way to start.
What it won't do
- Operate equipment. Claude for Manufacturing drafts documents. Machines are a separate project, the Model Hardware Standard, which is a research preview with a first group of labs and advanced manufacturers.
- Replace your ERP or quoting software. They stay the system of record. Claude works from their records and exports.
- Fix your data. It will tell you, honestly and at the top of every report, where your data is missing, stale or doesn't join. Closing those gaps is still your work.
The data work it assumes
Read the plugin closely and a pattern shows up. Almost every skill assumes things most mid-sized plants don't have yet: exports that land every night, identifiers that join across systems, an ERP it can reach, and the rules for which job runs first written down somewhere instead of living in one person's head. The plugin is honest about all of it. It reports the gaps instead of guessing around them. Someone still has to close them.
Most plants are in that position. In the Manufacturing Leadership Council's 2025 survey, 79% of manufacturers said manually entered data was among their data sources. 46% named data from disparate sources as a persistent challenge, and 45% the ability to extract data from legacy systems. Only 39% had a process to check that their data is accurate.
That gap is the work Databasin does. Our team connects the sources, prepares them in one lakehouse and keeps them running, so a tool like this has something dependable to read. For Claude for Manufacturing, that means:
- A connection to every system the plugin can't reach. Databasin's connector catalog covers the databases older ERPs and MES run on, including SQL Server, Oracle, DB2, SAP HANA, Sybase and Microsoft Access, plus file drops over SFTP, SMB shares and SharePoint, and vendor portals you can only download reports from.
- One prepared copy of the plant. Each source lands in a lakehouse, and the crosswalk and shared definitions (what counts as scrap, which downtime codes are planned) live there as maintained views. A private deployment can keep the platform, storage and model route inside your own Azure environment.
- Exports that someone watches. Each source loads on a schedule, and a failed run notifies a named person by email instead of leaving a gap for the morning report to find. Our team sets up the nightly files the plugin reads from that prepared copy, so every skill works from the same joined records.
- Your first chain running with your team. We adapt the skills to your own forms, templates and procedures, which Anthropic expects every plant to do, and test the chain next to the old way before anything changes.
The same prepared data works without the plugin, too. Databasin One runs on Claude against your governed views with read-only query tools, and shows the query behind each answer. Its LLM Agent task, in beta, runs a saved skill on a schedule and delivers the result, which is the same shape as the morning plant pulse.
Questions we expect to hear
What is Claude for Manufacturing?
A plugin for Claude, released by Anthropic on 12 November 2026, with 17 skills for quoting, quality, operations, supply chain and design. It drafts quotes, NCRs, 8Ds, supplier requests and production reports from a plant's own records, and a person reviews each one before it goes anywhere.
Which Claude plans include it?
Paid plans. Anthropic's announcement names Team and Enterprise, and Enterprise bought through AWS Marketplace is included. An admin installs it from the plugin directory.
Does it work with SAP, Epicor, Plex or JobBOSS?
Not out of the box. The ERP connectors are Oracle NetSuite and Infor M3. Any other system reaches the plugin through its custom connector skill or a scheduled file export, and eleven of its twelve single-job skills work from exports alone.
Do we need connectors at all?
No. Only design space explore needs a live connection, to Autodesk Fusion on the desktop. Everything else can run from a folder of nightly exports. Connectors make the data fresher, not possible.
Our ERP has no API. Can we still use it?
Yes, through exports. If the ERP keeps its records in a SQL database you can reach, a scheduled read-only query against that database is usually the most reliable way to produce them. Databasin connects to SQL Server, Oracle, DB2, SAP HANA, Sybase and Microsoft Access, among others. A system with neither can still send files over SFTP, an SMB share or SharePoint.
Does it operate machines?
No. It doesn't run production control, operate equipment or write to your systems of record. Anthropic's Model Hardware Standard, a separate research preview, covers how AI agents could work with equipment.
We're a process manufacturer. Is it for us?
Partly, today. Anthropic says the quoting and quality skills fit best where each job gets its own quote, drawing and routing. The morning report, breakdown diagnosis and supplier skills work for food, CPG, chemicals and steel now, and the rest can be adapted to batch records.
How long does it take to set up?
The “set us up” conversation fits in one sitting. How long the rest takes depends on your systems more than on the plugin: a plant on NetSuite with clean exports is close to ready, and a plant on a decade-old on-prem ERP has export and crosswalk work to do first.
Sources
Anthropic, Claude for Manufacturing, 12 November 2026 · claude.com/resources/articles/claude-for-manufacturing
Anthropic, Claude for manufacturing solutions, Accessed November 2026 · claude.com/solutions/manufacturing
Anthropic, claude-for-manufacturing plugin source, Accessed November 2026 · github.com/anthropics/claude-plugins-official
Infor, Infor plugin for Claude, Accessed November 2026 · claude.ai/directory/infor
Anthropic, Model Hardware Standard research preview, 2026 · anthropic.com/news/model-hardware-standard-research-preview
Manufacturing Leadership Council, 2025 survey · manufacturingleadershipcouncil.com
Databasin, Connector catalog, 2026 · databasin.ai/docs/connectors/catalog
Get Claude for Manufacturing working on your data
Tell us the process you'd start with and the systems you run. We'll walk through what it takes to connect them, build the crosswalk and keep the exports running, or tell you it's too early to start. We get your first use case live in 90 days.