AI Studio

A visual revision workspace

AI Studio is a Windows, shell-owned visual editing workspace for the application currently on screen. It maps the mounted interface, asks configured local models for bounded changes and previews the valid changes before you decide what to keep. Open it with Ctrl+Shift+A in a package that includes AI Studio.

Use it to explore spacing, hierarchy, typography and consistency across existing controls. It is not a repository editor, a background application builder or a way to give a model unrestricted access to your computer.

Before you start

Include uk.capp.sidecar.ai-studio in the application's Sidecar selection and produce an updated Windows package. An ordinary browser preview does not supply the required native model bridge. The declared target is Windows; a planned target is not a working native implementation.

Have an LM Studio-compatible local endpoint available on the same computer. The default endpoint is http://127.0.0.1:1234; AI Studio's host restricts this connection to loopback. Changing it to a remote provider is not a supported substitute. The browser page does not contact the model endpoint directly.

Configure Planning and Execution models. Vision is optional and needs a model whose catalogue entry supports images, unless you deliberately use the operator override. These are AI Studio's model-role settings, not the Local AI Sidecar's device-model settings. Selecting Local AI does not configure this pipeline for you.

Use a test account and non-sensitive application data. App-only capture excludes the surrounding cApp tools, but a screenshot can still contain information your application displays. Keep sensitive-text redaction enabled and inspect the selected context before sending it.

First successful revision

1. Open the application route you want to revise, let its controls finish rendering, and open AI Studio.
2. Set the local endpoint and each required role's model. Inspect the role rail for the actual selected model and loading, ready or error state. A saved model name alone is not a successful connection.
3. Refresh the UI map. Start with a bounded request such as “Make the spacing between these existing form controls consistent; preserve their labels and behaviour.”
4. Start the revision and watch the current focus and role handoffs. One focus is processed against the captured targets; the model does not get to invent arbitrary host actions.
5. Review the preview and the effective operations, including the unchanged parts of the interface. Reject unwanted revisions and accept only the result you have inspected.
6. Reopen the relevant route and test the original interactions, keyboard access and narrow layout. A visually plausible preview does not establish functional correctness.

Success means an actual effective visual change that you choose to retain. A model response, a completed focus or a favourable visual critique is not sufficient. No changes produced is not an accepted revision.

Models, budgets and handoffs

A role does not silently borrow another role's model. Model, context and concurrency settings belong to each role. The default routing keeps one model resident, verifies a matching resident instance or switches it before the next stage. Model loading and switching can therefore be visible parts of the workflow.

The supplied role-context defaults are not recommendations for every computer. Check the endpoint's model capability, supported context and observed memory use. A large context setting does not create additional device capacity. Model-role settings here do not certify GPU execution.

Structured results are locally checked. Malformed output can receive bounded repair attempts; invalid targets, invalid CSS and changes that do not alter the current computed value are not useful operations. Repeatedly asking the same failing focus to continue is not a substitute for resolving the reported model, transport or target problem.

Preview, persistence and recovery

Accepted sessions are app-scoped capp.ai-visual-profile.v2 patches. With profile persistence and automatic application enabled, a saved profile can affect subsequent views in that application. Acceptance does not rewrite the original project source, compile a new release or update already distributed binaries. Use the normal build workflow for source and package changes.

Keep review/approval enabled while qualifying a new model or application. A live preview is a candidate, not final acceptance. Cancelling active and queued work stops the current proposal flow; inspect the retained preview/profile separately rather than assuming cancellation erased previously accepted changes.

Transport failure pauses the current focus instead of promising a safe automatic replay. A completed session is for review, discard or acceptance, not endless resumption. When the application has navigated or replaced its DOM, refresh the map before making a new proposal against it.

Not the Local AI SDK

There is no portable AI Studio application API for submitting arbitrary code, choosing filesystem paths or running shell commands. Use the editing workspace's own controls; do not bind your product to its private controller objects.

Read/action settings do not turn model text into authority. The visual execution path remains bounded to validated visual operations; application actions and project operations have their own contracts. For product inference, use the inference API and a scoped Local AI connection instead.

Troubleshooting

Keep diagnostic exports private until reviewed: prompts, app context and model responses can contain application data even when credential fields are redacted.

Configuration reference

These are the selected package's configuration fields. String credential values and installation-specific defaults are never published here. A field's presence does not bypass its runtime policy.