Docs Designing the system 14
Designing with an AI
The AI drafts and assists; it never reviews or approves, the copilot never submits, and neither ever sees a secret's value.




- ✦ Copilot (a tab on every design: objects, services, connections, transactions, screens, flow
templates, report layouts, queries and menus): ask in words ("add a batch
reference field that quality fills in while the lot is on hold"). It reads the model, drafts into
your change, writes scripts with their test cases, dry-runs and validates them, as you, with your
rights. It can also bring another live design into the change, show what a role may do, run the
fitness test, check, lay out and explain a flow, walk an input flow, draft translations, and read a
service's runs in the integration monitor (what failed, and the line it failed at) before fixing it. What it drafted is marked Drafted with AI for your reviewers. You check and submit.
Ask it how the plant fits together. It reads what is live a part at a time, however large the plant: the
map of every design, one design whole (an object with its rule scripts), what a design uses and what uses
it, every place a field or a word is written, and the departments, groups and people. Ask "which
transactions touch a lot, and what rules apply?": it lists each transaction with what it does to the lot,
its checks, its signature, the other records it writes and the policies that let it, the lot's rule pipe,
and the objects a lot depends on (a reference, a rule that reads another object, a transaction or a route
that names both), two or three steps out if you ask. Before it changes a design it reads what the change
reaches, and tells you what else it touches. Each thing it read is a line in the conversation.
The conversation is kept with the change: come back any time (another day, after a restart) and
carry on where you left off. New conversation deletes it, after asking. Attach (or paste, or
drop, into the box) gives it what
you design from: a photo of a paper form, a drawing, a work instruction as a PDF, a list of values as
a CSV file or workbook; it says what it took from each, and designs only through your change, as
always. (A plan's input screen can ask a person for a file or a picture too: inputs of type
fileorimage, kept with the run.) It works on the change it was opened on and no other, whatever it reads there. Its saves build on the draft as it last read it: if a co-designer saved meanwhile, it reads theirs first rather than writing over it. A dry run it makes reads with your rights only: for a service that runs as a service role or as another caller, dry-run it yourself. There is a ceiling on messages (30 an hour for one person, 300 for the whole installation); past it, it asks you to come back later. - Which copilot. The three copilots (yours here, the integration monitor's, AI Report's) are one package that IT upgrades on its own, without restarting the plant's server: Designer → Suites → Copilot shows the version in use. When IT takes up a newer one, a copilot you have open switches to it by itself, keeping what you typed and the conversation; one answering at that moment finishes first.
- AI access (designer home): Issue token lets an outside AI (Claude Code, or another) work in
the designer as you, over MCP (
/ai/v1/mcp: Claude Code, Claude Desktop or any agent that speaks the Model Context Protocol, given the same tools as the copilot) or the REST API (/ai/v1, described by its OpenAPI). An installed suite may give the AI tools of its own, beside the designer's: named after it (<suite>__<tool>), they read or draft what only that suite knows, as you, over MCP and in the copilot alike, and go with the suite if it is removed. The AI reads which there are inget_suites(each with what it does and whether it reads or drafts); a suite's own guide says what its tools are for. Like every tool, none of them submits, reviews or approves. The new token is shown once, with the command that adds it to Claude Code (claude mcp add --transport http opencore-mes <your MES>/ai/v1/mcp --header "Authorization: Bearer <token>"): run it in a terminal, and Claude Code reads and drafts in your designer, as you. What an outside AI is given is kept: a tool, what it takes and what it answers stay as they are while/ai/v1runs (new ones may be added), and a tool due to go says so in its own description at least 180 days ahead, with what to use instead, so the AI moves without anyone changing it. Choose its scopes:design:read,design:draft, anddesign:submitif it may submit for you (the change's history then says the submit came through that token). There is no scope to review or approve. An outside AI reads the plant the copilot's way: the map (POST /ai/v1/index), one design (GET /ai/v1/designs/{kind}/{name}), what it touches (POST /ai/v1/related), where a word is written (POST /ai/v1/search), the people (POST /ai/v1/organization) and the run history (POST /ai/v1/runs). A token lasts 30 days, 90 (the default) or a year, then stops working; Revoke ends it sooner; a token also stops working the moment the designer is no longer shared with you. An AI that saves a draft should send thedraft_revit read (seen): a save built on a draft someone has changed since is then refused instead of undoing their work.