A second pair of eyes on your design

AI Architect reads your diagram and tells you where the design is weak: a database with no replica, a chain of calls that will be slow, a service that has no way to recover. Each finding comes with a reason, and many come with a fix you can preview and apply in one step.

What it does

Open the AI Architect panel in the editor. There are three ways to look at a design.

• Quick checks: free on every plan. A set of rules runs instantly in your browser and flags common problems, such as a database with a single copy, a long chain of synchronous calls, a database shared by several services, an event nobody consumes or publishes, and a public service reached without a gateway. They send nothing anywhere.
• Review: an AI reads the whole diagram and writes up what it finds.
• Deep review: four focused passes over the diagram, merged into one report. It takes longer and costs more.

The panel shows the price before you run a Review or Deep review. Findings are grouped into seven categories.

• Single points of failure
• Scaling
• Bottlenecks
• Data and consistency
• Resilience
• Security basics (for a full threat list, use the threat model)
• Patterns

You need edit access to a diagram to run a review. Anyone who can open the diagram can read the latest one.

Reading a review

Every finding has a severity (critical, high, medium or low), a short title and an explanation. Chips on the finding name the components, arrows and events it is about. Click a chip to jump to that item on the canvas.

Fixes

When a finding has a fix, open it with Preview fix. The preview shows the change as a readable list, such as a new component and the arrows that go with it, and highlights the items it touches. Nothing changes on the diagram yet. Check the change, then choose Apply inside the preview.

• Apply makes the change as one undo step. Undo takes all of it back.
• Apply is turned off if the items the fix needs have since been removed.

Fixes only change the Architecture tab: components, arrows and the consumers of an event. A fix never deletes a component that owns other items, such as a database with tables or a service with state machines.

You can add a short change note to the Notes tab when you apply. The note is separate from the fix, so Undo does not remove it. Dismiss hides a finding for the rest of your session. It does not delete it.

One review per diagram

Only the latest review is kept with the diagram, and everyone with access sees the same one. When the diagram has changed since the review ran, a banner says so, so you know the findings may be out of date.

Export

Export the review as Markdown, or choose Save as PDF, which opens your browser's print dialog on a print-ready report.

Chat

Under the report there is a chat. Ask why a finding matters, what an alternative would look like, or which risk to tackle first. The chat knows your diagram and the latest review. Each message costs 1 credit to 6 credits, by the size of the diagram, as a Review is. Chat is in the web app only. An agent uses review_diagram instead.

Credits

Quick checks are always free. Reviews and chat messages use credits. Credits are given to each person every month and reset on the 1st of the month (UTC).

AI Architect is part of Pro, Team and Enterprise. The free 14-day Pro trial includes 30 credits a month while it runs. Free has no AI credits; Quick checks keep working.

The credits are:

People in an organization get the credits of the organization's plan.

Reviews and chat messages are priced by the size of the diagram. A diagram that is too large to review is refused before it runs, with a message to split it or review one level, and costs nothing.

If a run fails, the credits are refunded, and a failed review shows no findings. A run keeps going when you close the panel or your agent stops waiting: it is saved and charged when it finishes. Only one review can run at a time for each person, so a double click never charges twice.

From your agent

The MCP server runs AI Architect with review_diagram, get_review and get_ai_usage, in the analysis toolset.

• review_diagram runs a Review or, with type: "deep", a Deep review, and returns the Quick checks, the findings and the credits left. Pass focus to point it at one part, such as the database layer. It costs the same credits as the web app.
• get_review reads the latest saved review, free. If a review takes longer than the agent waits, it still finishes and is saved, and get_review reads it without paying again.
• get_ai_usage shows the credits left this month, when they reset, what was spent, and the price of each review by diagram size. An agent should check it before a Deep review.

Fixes in an agent review are lists of edits. The agent is told to ask you before it applies them with the editing tools.

Privacy

• What is sent: nothing, until someone runs a Review, a Deep review or a chat message. Then the diagram content needed for that run is sent.
• Where it runs: only on Cloudflare Workers AI. The models run on Cloudflare. No data goes to the model makers (OpenAI, NVIDIA or Google).
• Training: Cloudflare does not train on customer data.
• What is stored: the latest review is stored with the diagram, so the people who can open the diagram can read it. A new review replaces it.
• Organization switch: organization admins can turn AI Architect off for the organization's diagrams. Members then get Quick checks only on those diagrams, their content is not sent to the model, and the agent tools are refused for them.

See the privacy policy for the rest of how we handle your data.