Windchill integration for AI agents: a Windchill MCP server that answers inside your contexts and config specs

For manufacturers running PTC Windchill who want Claude, or any agent that speaks MCP, to answer where-used, BOM and change questions from the system of record, without a shared super-user account.

Windchill integration for Claude and other agents works through an MCP server built on Windchill REST Services, the OData interface PTC ships. On 19 September 2026 PTC had no MCP server for Windchill. Pexon builds one with named tools such as where_used and compare_bom, per-user OAuth, explicit configuration specs and an audit log.

Checked 19 September 2026

What PTC and the community ship for Windchill and agents today

What existsWhat it doesWhat it does not do for an outside agent
Windchill REST Services (WRS)PTC's OData v4 interface to Windchill: parts, documents, changes, workflow. Supports OAuth with the authorization code flow.It is an API, not a tool list. Every question an agent can ask and every right it has must still be designed on top.
PTC statement on MCPIn February 2026 PTC wrote that Codebeamer and Windchill are evaluating their roadmap plans for MCP integration.No product, no date, no scope. Plan without it and re-check quarterly.
Windchill AI AssistantPTC's in-product assistant, launched April 2026, for working with Windchill content inside Windchill.It stays inside Windchill. It does not put a BOM next to the SAP material master.
Onshape FeatureScript MCP serverAn MCP server from PTC for FeatureScript in Onshape, announced August 2026.Onshape, not Windchill. It shows PTC ships MCP, not that Windchill has it.
Community Windchill MCP serversSeveral on GitHub. The most active describes itself as a test bed with 64+ tools for Windchill 13.0.2.x.One account's credentials for every user, and write tools beside read tools. A way to learn the API, not a production shape.

The gap is not access to Windchill: WRS is a documented, versioned interface. The gap is a server that knows who is asking, which configuration the answer is valid for, and which operations are allowed at all.

Why Windchill is hard for agents

Four places a naive Windchill connector gives a wrong or unauthorised answer

These are our assessment from the Windchill data and permission model, not war stories. Each one is a design decision the blueprint settles before code is written.

01 · Access

Context ACLs flatten under a service account

Windchill grants rights per product or library context, team role, domain and lifecycle state. A server that logs in as one technical user answers with that user's rights, so an engineer on product A sees product B's restricted parts through the assistant.

02 · Configuration

The wrong config spec returns a real but wrong BOM

Latest iteration, latest released and as-stored are three different structures of the same assembly. A BOM call that leaves the navigation criteria at the default gives an answer that looks right and describes a product nobody builds.

03 · Data model

Soft types and custom attributes differ per installation

The attributes your engineers search by are often soft attributes added over years of customisation. A generic server does not know them, and some may need a WRS domain extension before an agent can read them at all.

04 · Licensing

Agent access may be a contract question

Whether calls from an agent on behalf of named users need additional Windchill licences depends on your PTC agreement. We do not guess at it; the blueprint lists it as a question for your PTC account team.

  • Per-user identity, an explicit config spec per tool and a mapped attribute list are the three things the first server must get right.

The first tool list

Eight named operations for release one, all read-only

This is the artefact the blueprint produces and the PLM owner signs. Names and fields are adapted to your instance; the shape stays. No tool takes free OData or runs as anyone but the person asking.

# windchill-mcp, release 1: read-only, per-user identity
# Every call runs with the asking user's OAuth token (WRS authorization code flow).
identity: per_user            # no shared service account
config_spec: explicit         # every structure tool states which one it uses
writes: none                  # drafts for change requests come in release 2

tools:
  search_parts(query, context?)                 # number, name, version, state, context
  get_bom(part, config_spec, levels<=3)         # structure as this user may see it
  where_used(part, config_spec)                 # parent assemblies, per product context
  compare_bom(part, from_version, to_version)   # added, removed, quantity changed
  get_open_changes(part)                        # problem reports, change requests, notices
  get_change_impact(change_number)              # affected items, their parents, documents
  get_documents(part, type?)                    # related documents, metadata only
  compare_to_sap(part, plant)                   # eBOM vs SAP BOM delta, via the SAP server

audit: user, tool, arguments, result_count, duration  # one line per call

If a tool here would need a right the asking user does not have in Windchill, the call fails with Windchill's own error. That failure is the test that the identity model works.

What we build

The part no vendor assistant covers: Windchill next to SAP

A question about a part is rarely a Windchill-only question. Where is it used, is the change released, and does the SAP BOM for plant 1000 already reflect it: the last half lives in the ERP, and an in-product assistant cannot see it. The server we build publishes Windchill tools and, next to them, the few SAP operations the comparison needs, so the agent can put the engineering BOM and the manufacturing BOM side by side and name the delta.

Underneath the tool list sit the same three pieces every connector needs. The identity exchange from your directory to a Windchill OAuth token for that user. An audit log of every call with the user attached. And an evaluation fixture: twenty or thirty real questions with known answers from your data, run on every release so a Windchill upgrade that changes a response shape is caught before a user sees it.

It is your code, in your repository, deployed in your environment next to Windchill. Nothing routes PLM data through us.

What we need from you, and what happens after the pilot

  • A Windchill test or staging server on the same release as production, and one person from PLM administration for about a day a week during the blueprint.
  • The OAuth set-up for WRS, or the decision to set it up. Without per-user tokens the tool list shrinks to what everyone may see, and we write that down.
  • The list of soft attributes your engineers actually search by. It is usually shorter than the type manager suggests.
  • Your answer on licensing from PTC before the pilot starts, so the cost of agent access is known and not discovered.
  • After the pilot you run the server yourself with the runbook and tests we hand over, or we operate it; either way the next system reuses the identity and audit work.
  • If the server gives a wrong answer, the audit log shows which tool, which config spec and which user produced it. A read-only first release means a wrong answer cannot become a wrong change.

Start with the Readiness Blueprint

Two weeks, €4,900 fixed price. We take your Windchill instance, agree the first tool list with the people who own the data, prove on staging that a call runs as the asking user and respects their contexts, and hand back a costed build plan for the server. A first production server usually follows as a use-case pilot from €15,000. The blueprint is yours whoever builds the server. All prices are net and exclude VAT.

We are the wrong choice if you only want engineers to search Windchill documents inside Windchill. PTC's own assistant is built for that, and a separate server would add a system to operate for no gain.

Related

Questions about a Windchill MCP server

Does PTC offer an MCP server for Windchill?

Not as of 19 September 2026. In February 2026 a PTC representative wrote in the PTC community that Codebeamer and Windchill are evaluating their roadmap plans for MCP integration, and pointed customers to their account team. PTC has shipped an MCP server for Onshape FeatureScript and an in-product Windchill AI Assistant, but neither gives an outside agent access to Windchill data. We will re-check this page quarterly.

Can the agent see parts a user is not allowed to open in Windchill?

No, if the server is built the way we build it. Windchill REST Services supports the OAuth authorization code flow, so each call can run with a token issued for the person asking, and Windchill applies its own context, team and lifecycle rules to that person. Where your installation only allows a shared technical user, we cut the tool list down to what everyone may see and write that subset into the blueprint.

Why not use one of the open-source Windchill MCP servers?

Use one to learn, not to run production. The most active one we read calls itself a test bed, logs in with one username and password from its environment, and ships create, revise, check-in and lifecycle-state tools next to the read tools. That is a fine way to explore the API and the wrong shape for a plant: every user gets the rights of one account, and nothing distinguishes a question from a change.

What does a Windchill MCP server cost?

The two-week Readiness Blueprint is fixed at €4,900 and returns the agreed tool list, the identity test and a costed build plan. A first production server is usually a use-case pilot from €15,000. Whether API access by an agent needs additional Windchill licences depends on your PTC contract, so check that with PTC before the pilot; we list the question in the blueprint. All prices are net and exclude VAT.

What will a Windchill MCP server not do?

It will not change released data on its own. The first release is read-only; writes come later and only as drafts a person submits. It does not interpret CAD geometry, only metadata and structure. Heavily customised soft types can add build time we only see in the blueprint. And we have not yet delivered a Windchill server for a named customer, so the plan is built from the API and your instance, not from a previous project.

Next step

Bring one Windchill question you would like an agent to answer.

Thirty minutes with the architect who would run the blueprint. We look at your Windchill release, your authentication and the first five tools, and tell you whether a server is worth building.