MCP for Historians
Give your AI assistant a safe, read-only window into your historian
A read-only Model Context Protocol (MCP) server that connects Claude, ChatGPT, and Cursor to Canary Historian™ or AVEVA™ PI System, with analytics that return answers instead of raw data.
Example conversation. Tag names and values are illustrative.
Ask in plain English
Ask your historian questions from the AI assistant you already use. No SQL, no exports.
Works with your assistant
Claude, ChatGPT, Cursor, Codex, or any other MCP client, so your team keeps the tools they already know.
Built for process data
Understands tags, asset models, alarms, batches, and quality codes, not just generic tables.
One toolset, two historians
The same prompts work on Canary and AVEVA PI, so sites on different historians ask questions the same way.
Live data, no copies
Answers come straight from your historian, with no exports or extra copies of your data to manage.
Guided workflows
Ready-made prompts for a morning summary, a comparison with yesterday, or triaging a failing tag.
Plant documents, too
An optional add-on looks up P&IDs, procedures, and datasheets alongside the process data.
Read-only by design
An AI assistant connected to production data should never be able to change it. This one has no way to.
No write path
No tool can write a value, change historian configuration, or acknowledge an alarm.
Your permissions apply
On Canary, each person signs in with their own token. On AVEVA PI, reads run as a service account you control.
Role-based access
Read-only, operator, and admin roles control who sees live data and diagnostics, with no restart to change them.
Single sign-on
Sign in through Microsoft Entra™, Okta®, or any OpenID Connect provider, and offboard a person in one step.
Fully audited
Every call is logged with who made it and which tool ran. Tokens and values are never logged.
Guardrails on every query
Rate limits and time-window caps stop a runaway question from loading the historian.
Fewer tokens, lower AI cost
With skill files alone, the assistant pulls raw data into its context and does the math itself. This server does the math and sends back the answer.
Answers, not raw series
Ranking 40 compressors comes back as about 80 numbers, not 40 raw time series.
Math on the server
Averages, baselines, and statistics run next to the historian instead of inside the AI's prompt.
Guidance on demand
Reference docs load only when a question needs them, not at the start of every conversation.
No scripts to write
The assistant calls a ready-made tool instead of writing, running, and debugging its own code.
Answers, not data dumps
Analytics run on the server, so the assistant gets a result instead of thousands of raw points.
Computed verdicts
Baselines and statistical tests return normal, elevated, unusual, or extreme instead of a spreadsheet.
Fleet rankings
Compare one metric across dozens of assets in a single question.
Change detection
Find exactly when a signal shifted instead of eyeballing a trend.
Scorecard reports
Define a recurring report once, then run it against fresh data any time.
Gaps are reported
"Nothing happened" and "couldn't look" are always different answers, so you can trust a clean result.
Troubleshooting tools
Ask why a tag is returning nothing and get a diagnosis with the most likely cause.
Deploy next to your data
Runs on Windows inside your network, close to the historian.
Simple architecture
One self-contained Windows executable for remote HTTP MCP connections.
Guided install
A signed setup wizard installs the service, certificate, and firewall rule for you.
Operator portal
Manage settings, user access, activity, and licensing from a browser, with every change backed up.
Private by default
Keep it entirely on your private network, with each user connecting through a small local proxy.
Reach from anywhere
Publish over HTTPS or a Cloudflare® Tunnel when people need claude.ai or ChatGPT access.
Offline licensing
One licence per server, checked on the machine itself. Nothing calls home.
Supported historians
Each server connects to one historian. Running both? Install one for each, and your assistant can use them side by side.
Canary
Tag discovery, history, events, asset models, and fleet analytics on Canary Historian.
Model-aware tag search
Search every view at once, with each result flagged for whether it actually holds data.
Hierarchy navigation
Walk datasets, nodes, and tags, even in very large hierarchies.
History & events
Read historic values, alarms, operator annotations, and batch or campaign events.
Asset model
Query asset types and instances to compare like equipment with like.
Condition scans
Find every time a condition held, like discharge pressure above 75, across a whole asset type.
Historian health
Check storage, archiving, and collector connections from Canary's own diagnostics.
Live data
Poll current values for a chosen set of tags, for operators and above.
Built-in product knowledge
An embedded Canary knowledge base answers configuration and troubleshooting questions too.
Requires Canary System 25.4 or later. Each person signs in with their own Canary API token, so the server never shows anyone more than they can already see in Canary.
AVEVA PI System
AVEVA PI Data Archive points, AF hierarchies, event frames, and AVEVA PI Vision displays.
Points and attributes
Search PI points and AF attributes by keyword and see their units and data span.
AF navigation
Browse AF databases and elements, and compare elements built from the same template.
History & summaries
Read archived values or PI's own summaries, with quality and system states decoded.
Event frames & batches
Read event frames by template and severity, with the templates you choose shown as batches.
Calc Failed, explained
Trace a bad AF attribute back to the analysis that writes it and the most likely cause.
AVEVA PI Vision displays
Find which displays use a point, and check whether their bindings still resolve.
Live data
Poll current values for a chosen set of points or attributes, for operators and above.
Scoped to your AF databases
Choose which AF databases users can reach. Everything else stays hidden.
Reads through the AVEVA PI AF SDK, so AVEVA PI Web API is not required. Reads run as the service’s Windows account, so your existing PI security for that account controls what the server can see. Users sign in with personal access tokens from the operator portal, so the audit log names each person, or through your identity provider.
AVEVA PI Vision tools need AVEVA PI Vision 2021 or later.
Ready to put AI to work on your process data?
Tell us whether you run Canary Historian, AVEVA PI System, or both, and we'll set you up with a pilot. Don't see your historian listed? Contact us and let's discuss developing an MCP server that meets your needs.