Troubleshooting

Sign-in

"Sign in with Google" does nothing

Most often a third-party-cookie or popup-blocker issue.

• Allow popups for prompt2cad.com in your browser settings.
• Allow third-party cookies for accounts.google.com.
• If you're in incognito / private mode with strict tracking protection, switch to a normal window.
• If Safari is showing "Prevent Cross-Site Tracking," temporarily disable it for this domain.

Google chooses the wrong account

The chooser may default to a previously-used account. Sign out from accounts.google.com first, then retry.

"This site can't be reached" after Google consent

A redirect URL mismatch. This is a server-side configuration issue — email support with a screenshot of the URL bar.

Billing

“Account temporarily suspended”

This notice appears when you open the app or a project while your account access is temporarily restricted. Account access is checked as part of sign-in/session loading; active accounts do not see a separate account-check loading screen. Generation, project access, exports, and new credit purchases are unavailable during suspension. Suspension itself does not delete your saved projects or reset your credit balance, and it is not a refund.

If the notice mentions payment verification, Stripe requires identity verification related to your payment. Follow any verification instructions you received from Stripe. For help with your Prompt2CAD account or a review, use Contact support or email support@prompt2cad.com.

After support confirms the issue is resolved, select Check again. The app also checks automatically every 30 seconds and when you return to its tab. Rechecking is free; it does not purchase credits or override an unresolved restriction. You can still sign out and contact support. Sign-in, a new payment, or refreshing the page does not by itself lift a suspension.

“Account access could not be checked”

The service could not verify your account status. This message does not mean your account is suspended. Select Check again; contact support if it persists.

Top-up succeeded in Stripe but my balance didn't update

The Stripe webhook lands in 1–5 seconds normally. If 30 seconds have passed:

• Hard-refresh /app (⌘/Ctrl + Shift + R).
• Check the credit chip in the header — it polls separately from the chat.
• If still missing, email support with the Stripe receipt number so we can look into it.

"Insufficient balance" but I just topped up

Same root cause as above — the credit chip should refresh on its own; if it doesn't, a hard reload fixes it.

"Insufficient balance" with credits still remaining

Each AI call needs a temporary reservation for its input, response allowance, and any required conversation compaction. Chat automatically uses a smaller response allowance when a useful smaller call fits, but a positive balance can still be below the minimum hold. Top up to continue; this blocked call has not started or been charged. Earlier calls in the same generation may already have used credits. See Credits for the billing rules.

My card was declined

Stripe shows the decline reason on the checkout page. Most common:

• Insufficient funds (try a different card).
• 3D Secure / SCA challenge timed out (retry, complete the bank prompt promptly).
• Geo-restriction (your card's issuer blocks foreign currency).

If your bank says "the merchant declined," Stripe or the payment network may have blocked the transaction. Follow the Checkout guidance or contact Link support for transaction-level help; Prompt2CAD cannot override the payment decision.

Export paywall opens even though I've paid

The app verifies download access when you click export and fails closed if the check cannot complete. If you have paid but the export remains gated, check whether a reconciled refund or older charge left the balance negative. Otherwise:

• Hard-refresh.
• If still gated, sign out and back in — your session may have expired.

Geometry / viewport

Red error banner: "Code error: ..."

The geometry code threw at execution time. Almost always either:

• A parameter combination the agent didn't anticipate (e.g. you dragged width so low that a derived dimension went negative).
• A reference to a section / primitive that doesn't exist (rare — usually means the agent's edit lost track of a name).

Fixes, in order:

1. Drag the offending parameter back to its previous value.
2. Use Undo to revert the last edit.
3. Ask the agent — "there's a code error: <paste the message>" — and let it fix.

Viewport is empty / "Your model will appear here"

The current model has no geometry. Either:

• You started a new project and haven't prompted yet.
• The last build failed (check the red error banner).
• The agent cleared the model (rare, intentional reset).

Send a prompt to build something.

Sliders are sluggish

Models with very many primitives (1,000+) re-render slowly on parameter changes. Check the primitive count in the viewport's bottom-left chip. If it's high, ask the agent to simplify the model — for instance, by replacing decorative detail with a single shape, or by removing parts that are not load-bearing for the design.

Wireframe view shows missing geometry that's there in rendered view

Some primitives are made via CSG (boolean operations). Wireframe renders the post-CSG mesh, which can look messier than the rendered solid. This is cosmetic; exports still produce the rendered geometry.

Camera won't reset

Press R with the viewport focused. If you've panned far off the model, the camera "loses" the origin. R re-fits.

Chat

An independent review reports "Requests ending with a model turn are not supported"

This Gemini error can occur when continuing after screenshot results. The updated transport supplies an explicit continuation after multimodal tool results, so the reviewer can continue without repeating the design generation. Reload after the update before starting a new review. A previously failed review card remains a record of that attempt; it is not a completed verification.

Experimental server-side access

Server-side generation is enabled selectively for signed-in users. Eligible new text-only projects select it automatically; other users keep browser chat. Existing browser projects and new image/CAD attachment handoffs keep their current workflow. The runtime is saved per owner and project, so changing access does not move an existing project between browser and server, including when opened on another device. Normal generation credit charges still apply.

A previously confirmed browser project can reopen its cached chat even if the runtime availability check is temporarily unreachable. On a first visit with unknown ownership, saved history stays readable while chat waits to connect; the composer remains disabled and a queued starting prompt is retained until selection succeeds. Checkpoint Restore also waits for confirmed browser ownership; viewing and forking saved versions remain available. Parameter editing, file switching and automatic AI naming also wait until ownership is known; the cached model remains viewable during that check.

If server-side generation is temporarily disabled, a server-owned project can still display saved progress and accept Stop. New server-side work is blocked until availability returns; it does not automatically start another browser turn.

The preview chat stays on "Understanding your request" or its input stays disabled

In the experimental server-side preview, opening a project first checks the durable Agent's saved state. After a server restart or preview deployment, an interrupted generation resumes from its saved transcript and model through a background continuation. That recovery does not require sending another prompt; Stop still cancels it. A slow AI request may delay the next visible change.

If a status read takes more than 15 seconds, the chat shows a reconnect error and retries automatically. This read does not stop the Agent or start another generation, and checking status has no credit cost. If a fiber has already failed, the next status check surfaces the failure instead of keeping the progress indicator running. Share the project URL with support if it continues to show no progress; the label alone does not confirm that the model is actively generating.

After you answer a question or approve a plan, an interrupted Agent turn can rejoin its saved request automatically if its original waiter is no longer active. Your answers are retained and are not asked again. The continuation keeps the same segment budgets and replays an already-paid request when available; subsequent model work uses normal credits. A turn you stopped is not restarted automatically. Thinking, text, and tool activity are published while the turn runs, rather than waiting for the final reply.

"Project agent state limit exceeded" in the preview

The previous preview implementation stored CAD, conversation, and recovery data inside an 8 MB Agent-state budget. These larger payloads now live separately in chunked backend storage; the execution state contains small references instead. Projects can continue past 8 MB of accumulated CAD/chat data without that artificial failure. Existing projects migrate when opened, including paused or interrupted turns; migration itself does not regenerate geometry or use credits.

If a turn failed before this fix, refresh the project and submit your requested change once more. Retry generation uses normal token-based credits; refreshing or checking status does not. This change does not make storage or model context unlimited, change the first-upload limits, or move your browser checkpoints. If a storage error repeats, share the project URL with support.

Agent stopped mid-turn

The app automatically retries temporary connection failures and incomplete model responses a limited number of times. These retries resend the same request without adding messages to your conversation. Thinking/progress remains visible; the reply and its model-changing actions are accepted after a complete response arrives. Previously completed actions remain in place.

If the connection drops after a request has been charged, automatic recovery replays its saved result rather than generating and charging for a replacement. A compaction summary already received is reused when a new model attempt is needed. Recovery may briefly wait for the original request to finish.

When a model hits its per-response output limit, complete actions are applied once and the agent continues from the current conversation and saved design. If no complete action was returned, the app can make up to two automatic continuations before showing an output-limit notice. Partial text remains in the conversation; unfinished tool calls are not executed.

Automatic continuation uses normal AI credits and remains subject to balance, generation limits and Stop. It does not rerun the already-paid request or restart your model. If it still cannot finish, ask for a smaller next step.

Stop cancels both the active request and pending retries. If automatic recovery is exhausted, an error appears:

• Transient failures (an interrupted stream, provider overload, a rate limit) show a Try again button. It continues the generation from its last completed step — work already applied (geometry, plans) is kept, nothing half-finished from the failed attempt is kept, and it works for image-only prompts too.
• If the error concerns a connection or output limit, you can send a follow-up asking the agent to continue from the current model. For repeated output limits, ask for a smaller change.
• A context-size notice concerns the accumulated conversation; the agent can summarise older context. An output-limit notice concerns one model response and does not by itself mean that the conversation is full.

"Service temporarily unavailable"

The AI model is rate-limited, geo-blocked, or down. Retry in 30 seconds. If persistent, the provider's status page (currently status.cloud.google.com) shows outages.

Agent says it can't see something I just changed

Conversations auto-compact past ~600k tokens. Compaction summarises old messages — usually transparently, but occasionally a key detail gets summarised away. Re-state the fact in your next message.

Replying in the wrong language

The agent picks its reply language from your most recent prompt. If you started in English and switched to Italian, the next reply might still be English; one Italian message fixes it.

Exports

Export downloads a 0-byte file

Almost certainly a browser permission / popup issue — the download dialog was blocked. Check the browser's download bar or notifications, then re-export.

"Format not supported by my CAD tool"

Per-format notes:

• STEP — try a different version: if your tool reads "AP203" but not "AP214," that's a tool-side issue. We only emit AP214.
• DXF in FreeCAD — use the Mesh workbench, not Part (FreeCAD's Part workbench ignores 3DFACE).
• DAE in SketchUp Pro — STEP is the better choice. DAE is for SketchUp Free.
• OBJ in Blender 3.x — make sure both .obj and .mtl are unzipped into the same folder before importing.

Exported model is 1/1000 the expected size

Your downstream tool is reading the units wrong. STEP / DXF are millimeters; GLB / OBJ / DAE / STL are meters. The import dialog usually has a units override.

Render takes forever

Renders queue server-side and take 30–90 seconds normally. If you've been waiting over five minutes:

• Don't navigate away — refreshing the page during a render can leave you charged for a render that never completes.
• Check the chat for an error.
• If genuinely stuck, email support with the project name and we'll look at the project state to figure out what happened.

Projects

Project disappeared

Most common cause: you cleared browser data, and the project was signed-out-local (not synced to the cloud). Sorry — that data is gone.

If you were signed in: check the project manager. Local copies are removed on logout in some browsers, but the cloud copy should still be there. Sign in and the projects should reappear.

If a signed-in project really vanished, email support with the project name. We don't guarantee recovery — server-side data is best-effort, not a backup service — but we'll see what we can do.

Multiple tabs editing the same project

Don't. The browser-side storage is single-writer; two tabs writing the same project race, and one wins. Use checkpoints if you want to compare two states; use separate projects if you want truly parallel work.

Still stuck?

Email support@prompt2cad.com with:

• The project name (and a screenshot if it's a visual issue).
• The approximate time the issue happened (UTC is fine).
• What you expected vs what you saw.

We have basic usage telemetry — most issues can be diagnosed from the event log without you having to reproduce, though for complex bugs we may ask you for steps to reproduce.

For broader product questions (pricing, formats, what it can design), see the FAQ on the main site.