AI for systems engineering
AI for MBSE, with people in charge of the model.
In Holarch, the model is the engineering record. AI reads it and proposes changes to it. A person reviews each proposal, applies what is right, and can reverse it with one undo.
Human authority
AI proposes. People decide.
No AI feature and no connected AI tool changes the model on its own. The same rules apply to Holarch's own AI tools and to the proposals of AI clients such as Claude.
Previewed first
Every proposal opens in a preview: a word-by-word diff of each rewrite, a table of proposed functions or links, or verdicts with reasons. Nothing changes until you apply.
One undo step
An applied change is one undo step, however many items it touches. It is saved, shared with collaborators and recorded in version history like any edit.
Risky changes start unticked
Changes to approved, baselined or locked items, removed links and deletions, and text with a web address are highlighted and start unticked.
Replies are checked
A reply with the wrong shape or with ids that were not in the request is rejected. Rewritten requirement text is checked against the quality rules, and the preview shows the result.
In the lifecycle
Where AI helps in systems engineering work
Each AI tool sits on the page where the work is done and reads the items that work needs.
| Activity | What the AI does |
|---|---|
| Requirements | AI Quality Review judges Correct, Feasible and Necessary and finds ambiguous or unverifiable wording. Rewrite, make measurable, split compound requirements, draft rationale and generate child requirements. |
| Functional analysis | Extract functions from a requirements document, with the requirements each one answers, then suggest the flow of the FFBDFunctional Flow Block Diagram A diagram of functions in the order they run, with branches, loops and parallel paths.. |
| Traceability | In the Traceability Matrix, propose missing links and flag doubtful ones, each with a confidence and a reason. |
| Verification | Draft a test case for a requirement: steps, set-up, expected result and the verification method. |
| Risk and decisions | Draft risks for an item, suggest trade study criteria from requirements and MOEsMeasure of Effectiveness A measure of how well a system meets the mission need, from the user's point of view., and draft decision rationale. |
| Reviews | Write a project brief: the changes since the latest baseline, what is at risk and the next actions. Summarize or translate a document. |
Ask AI, grounded in Holarch
Ask AI answers questions about the open project and about Holarch. It uses the project's items and relationships, the current page, the Help and a list of the features in the running app, so it says when a feature does not exist. Answers link to the items and Help topics they use, and Ask AI changes nothing.
Model context
Structured access for Claude and other tools
AI tools work best with the model itself, not with an export. Holarch gives them classes, relationships and checks, and keeps the person who owns the model in the loop.
MCP for AI clients
- Add Holarch to Claude (web, Desktop, mobile or Claude Code) at https://app.holarch.app/mcp. Sign in and approve; no token to paste. Other MCP clients connect with a personal access token.
- The client reads projects, the schema, entities and documents, and runs the model checks and the requirement quality check.
- Its changes collect in a proposal. You review each change in Holarch, apply the ticked ones as one undo step, or reject the proposal.
- Choose read only or read and propose changes, all projects or chosen ones, and when the access ends. Disconnect ends it at once.
REST API for scripts
- Read and change projects from scripts, exports and dashboards with a personal access token.
- Tokens can be read only and limited to chosen projects, with an expiry date.
- Your project role still applies. Changes are recorded in the project's history under your name, like edits in the app.
Controls
Controls for organizations and classes
Turn AI off
Instructors and organization admins can turn AI off for every project of a class or organization, for example during an exam. Requests are refused on the server, also with members' own keys. API and MCP access can be turned off with it.
Restricted projects
A project with a classification, CUIControlled Unclassified Information US government information that is not classified but must be protected or controlled by law, regulation or policy. or export-control (ITARInternational Traffic in Arms Regulations US rules that control the export of defense articles, services and related technical data./EARExport Administration Regulations US rules that control the export of commercial and dual-use items and technology.) marking has AI turned off. Nothing is sent.
Only what a feature needs
Nothing goes to the AI provider until you start an AI feature, and then only what that feature needs. Holarch stores no prompts and no replies, and does not use your content to train AI models.
Marked data
Project text is sent to the AI as marked data, and the AI is told not to follow instructions found in it. A preview shows a banner when the input included imported or hidden text.
Limits per key
Set a monthly token limit on a key shared with an organization. Viewers cannot use AI tools.
Works without AI
AI is optional. Everything else in Holarch works with it turned off.
Turn AI off for an organization → What is sent to the AI provider →
Plans
AI included on every plan
Holarch's AI is Claude, from Anthropic. Every plan includes AI requests each month, so AI works without a key.
Free / Academic
75 requests a month with Claude Haiku.
Individual
150 requests a month with Claude Sonnet.
Team
300 requests a month per seat with Claude Sonnet, and Claude Opus for heavy tasks.
For more, add your own Anthropic key or use your organization's. Anthropic bills that account, and those requests do not count against the included ones. Keys are encrypted on the server and never sent back to the browser. Why Claude · Pricing