For operators in power, water, chemicals and process manufacturing
PI System integration for Claude means an MCP server on PI Web API that reads AF elements, event frames and summaries from the DMZ copy of the archive, as the person asking. AVEVA announced MCP integrations on 20 May 2026 without a date, and on 19 September 2026 no vendor server had shipped. Pexon builds one, read-only.
Why PI is hard for an agent
01 · Naming
Archive tag names follow whatever convention each site used when it was configured. An agent that searches them as strings answers confidently from the wrong tag. AF templates fix this, but only where the site has actually modelled its assets in AF.
02 · Zones
PI commonly runs in the DMZ at Purdue level 3.5, or as a PI-to-PI copy of a level 3 server. An agent in the office network has no business reaching the original, so the connector reads the replicated or enterprise copy and nothing else.
03 · Volume
A recorded-values call over a long window returns every stored event and loads the archive while it does. The connector has to force summaries and interpolated values with a capped point count, so a vague question cannot become a bulk export.
04 · Writes
PI is rarely a control path, but it is the record engineers and auditors rely on. An agent allowed to write values or edit AF attributes turns every wrong answer into a wrong record, which is why the tool list has no write operation at all.
Checked on 19 September 2026
| Option | What it is | What it leaves to you |
|---|---|---|
| AVEVA, on-premises PI | Announced on 20 May 2026: AVEVA "will bring Model Context Protocol (MCP) integrations" to on-premises and cloud systems. No ship date. | Everything until it ships. After that, anything that crosses from PI into another system, such as maintenance orders. |
| AVEVA CONNECT AI Assistant | An MCP server described on AVEVA's tech blog for the CONNECT assistant: metadata search, GraphQL query, context. | It serves CONNECT, the cloud platform. An on-premises PI Data Archive is not covered by it. |
| Build in-house | Mitsubishi Heavy Industries presented at AVEVA World 2026 an MCP for PI giving agents access to AF tags, time series and event frames. | Proof that it can be done on real assets. It was built in-house, with the maintenance that comes with that. |
| Community repositories | Small open-source servers on PI Web API, such as mingzilla/pi-api-mcp-server, last pushed in May 2025. | Hardening, identity, query budgets and maintenance. |
| US integrators | IACS Engineering offers a consultation on an MCP server for the factory historian, naming AVEVA PI, GE Proficy and AspenTech IP.21. | A real alternative. Ask where the server sits and whose identity each query runs under. |
What we build
01
The server talks to the PI Web API instance on the DMZ or enterprise side, over the authentication method your PI Web API is configured for. It never opens a path toward the level 3 archive, and placement is agreed with whoever owns the zone model before the first call.
02
Tools resolve assets through AF templates, return values with units, and answer time questions with summaries or capped interpolation. Where AF coverage is thin, we write down which assets the agent can see and which it cannot, rather than letting it guess from tag names.
03
Each call runs as the person asking, so that PI's own mappings and AF security decide what they see; the Blueprint tests whether your setup allows this before the build. Every call is logged with the identity, the tool, the arguments and the size of the result. It is your code in your repository, deployed in your environment.
The first tool list
This is the list a PI owner signs before anything is written. Each tool maps to PI Web API calls for elements, attributes, streams and event frames on the DMZ copy.
# First tool list: PI System MCP server (read-only, PI Web API, DMZ copy)
# Assets are found through AF templates, never by pattern-matching tag names.
find_assets(template, name_filter, root_path) -> AF elements with paths
get_asset_attributes(element_path, attribute_names) -> current values with units
get_summary(element_path, attribute, start, end, type) -> average | minimum | maximum | total | count
get_interpolated(element_path, attribute, start, end, interval) # point count capped per call
list_event_frames(element_path, template, start, end) -> trips, downtimes, batches
get_event_frame(event_frame_id) -> start, end, reason, attribute values
# Not in the list: raw recorded-value export, tag search by name pattern,
# manual data entry, and every PI Web API write method (POST, PUT, PATCH, DELETE).The omissions are the point. A reviewer should be able to read the last two lines and know what the agent cannot do without reading the server's code.
When to wait for AVEVA instead
If PI is the only system the agent needs, you are content with a cloud route through CONNECT, and AVEVA names a date that suits you, waiting is a reasonable choice. Build now when the question crosses systems, such as a trip in PI against the work order in maintenance, because no vendor server crosses its own product boundary. Why we keep even the vendor's own servers read-only is argued on industrial MCP server security.
Two weeks, €4,900 fixed price. We review your AF model with the PI owner, agree the read-only tool list, test whether PI's own permissions can apply per user through the connector, and hand back a costed build plan. What we need from you: read access to a PI Web API instance on the DMZ or a test copy, one person who knows the AF model, and one question the agent should answer. The plan is yours whoever builds it. All prices are net and exclude VAT.
No PI connector has been built for a named customer yet, and we will not suggest otherwise. The method is the same one we use for every MCP server; what is specific to PI is written on this page.
Not one you can deploy, as of 19 September 2026. AVEVA's press release of 20 May 2026 says it will bring MCP integrations to on-premises and cloud systems, without a date. An AVEVA tech-blog post describes an MCP server for the CONNECT Industrial AI Assistant, which is the cloud platform rather than an on-premises PI Data Archive. When AVEVA ships, the tool list and identity model we build carry over; the transport underneath changes.
PI Web API, in almost every case. It is REST and JSON, it covers AF elements, attributes, event frames and streams, and it can be called from any language the server is written in. The AF SDK is a .NET library and makes sense only where the server must run inside an existing .NET estate or needs an operation PI Web API does not expose. Either way the server reads the DMZ or enterprise copy, not the level 3 archive.
No. The first tool list is read-only and stays that way. PI is rarely a control path, but manual data entry and AF edits change the record that engineers, auditors and regulators rely on, and an agent that can write there turns every wrong answer into a wrong record. If a use case needs a value recorded, the agent drafts it and a person enters it.
The two-week Readiness Blueprint is fixed at €4,900: the AF model reviewed, the read-only tool list agreed with the PI owner, a test of whether PI's own permissions can apply per user, and a costed build plan. A first production connector is usually a use-case pilot from €15,000. All prices are net and exclude VAT.
It will not make a thin AF model rich. Where assets are not modelled in AF, the agent is left with tag names, and on most archives those were written for the people who configured them, not for anyone asking a question later. It is also not a control-room tool: answers arrive in seconds, from a replicated copy, and belong in analysis, handover and root cause work, never in anything that has to react in real time.
Next step
Thirty minutes with the engineer who would build the connector, with your AF model open if you can share it.