AI Copilot

Copilot is an AI assistant built into 360 Hextile. It answers questions about the app, changes your render settings for you, navigates you to the right screen, and can operate the draft in an open Layers tool. It knows your GPU, your installed models, your license, your current settings, and your recent renders — so its answers are about your setup, not a generic manual.

Open it with the chat bubble in the bottom-right corner. It's on every screen, not just the Dashboard. Hide it under Settings → AI Copilot → Show Chat Icon if you'd rather not see it.

What it can do

• Answer questions about features, pipelines, templates, and settings
• Change your settings — ask for "more detail in the sky" and it proposes the exact parameter changes, shows you what's changing, and applies them on your click
• Take you somewhere — "open my last render", "show me the LoRA batches", "where do I set the output size"
• Operate an open Layers draft — align, arrange, style, duplicate, remove, or run supported per-layer actions using the exact visible stack and selection
• Suggest prompts for a style or scene
• Troubleshoot render problems using your actual hardware and configuration

Copilot can use Generate Layer and Spot Clone only through the guarded flows below. It will not arm a tool for you, choose another render or clone source, or silently land or commit a result. It cannot start an ordinary image render or operate a closed tool. Prepare Clone from… on the Spot Selection card before asking Copilot to commit that exact live clone.

Setting it up

Copilot needs a language model. You have two options.

Your own API key (recommended). Paste a Claude, OpenAI, or xAI (Grok) key into the panel the first time you open it — Copilot walks you through it. There's no usage limit, no round-trip through our servers, and you pick the model. Keys are stored locally and encrypted.

Switch model any time from the selector in the panel header. The list follows your key's available models, not a frozen SKU sheet.

No key. With an activated license, Copilot falls back to our hosted assistant — no setup, but limited to 20 questions per hour. Add a key to lift the limit.

Keys are shared with AI Prompt Transform — set one up and both features work.

Changing settings

When you ask for a change, Copilot writes its answer, then shows a change card listing exactly what it wants to modify:

Steps        30 → 45
CFG Scale    7.5 → 6.0
Template     hexagonal_44 → hexagonal_100

Values already matching are greyed out as (already set). Expand Raw to see the underlying JSON. Then:

• Apply — merges the changes into your live settings
• Dismiss — hides the card
• Download — saves the config as a .hextile.json file instead of applying it

Only the listed fields change. Everything else in your configuration is untouched — Copilot merges, it never overwrites your setup with a fresh config.

Undo

After applying, the card turns into Applied ✓ with an Undo button, and an undo counter appears in the panel header. The snapshot lives with the message and survives restarting the app or switching conversations. Undo restores it only when Copilot can prove that the open document is the one that created the snapshot. If old or anonymous history cannot prove that match after a restart, Copilot refuses the restore instead of applying it to the wrong document. Changed your mind again? Redo puts it back while that same document match still holds.

Apply modes

The Off / Ask / Auto toggle in the panel header controls how eager Copilot is.

Auto is not enabled in the current build — the button appears greyed out. Ask mode is the intended day-to-day setting; it costs one click and you always see what changed first.

If Copilot proposes switching your pipeline, or if a change triggers a validation warning, it will ask even when it otherwise wouldn't. Pipeline switches reload the model and are never silent.

Working with Layers

Open Layers on the panorama you want to edit, then ask for one concrete action. Copilot can align and distribute selected layers; change pose, size, opacity, blend, visibility, lock, flip, or stack order; duplicate or remove layers; edit effects; and start supported upscale or cutout operations.

Ordinary Layers edits are immediate in every Apply mode because they act on the visible draft, not on the render-settings change card. Before each edit Copilot validates the exact armed tool, panorama, selection, stack, and draft revision it was shown. Generate Layer uses a separate guarded flow: arm Layers yourself, then approve generation, landing the result, and committing the draft as separate actions. Copilot never arms or retargets Layers for you.

The result appears in the conversation and in Layers' live Undo control. Undo reverses a live layer insertion; it does not cancel a generation job. After the draft is committed, use graph history to undo that graph change. The Copilot history is an audit trail, not a replay engine. Long-running upscale and cutout actions report progress and can be cancelled. Closing only the chat panel does not stop them, but a new request, a new conversation, leaving Layers, or changing the active panorama does.

Navigation

Ask Copilot to take you somewhere and it offers a link to that screen — a settings tab, your render library, a specific render, a template, a LoRA batch. It can only link to real places inside the app or to 360hextile.com; it can't be talked into opening anything else.

Automation: search saved prompts

Ask “Find saved render or sequence prompts mentioning sunset” to search the global and tile prompts stored with your renders and sequences. Searches need at least two characters; results give the saved item's kind, exact ID, and a short matching snippet. They do not search image pixels, captions, or other metadata.

Automation returns up to 50 hits per page. If a page includes next_cursor, continue with that cursor and the same query and filters until next_cursor is null; an empty page can still have a continuation. You can narrow a search to renders or sequences and to an active, archived, or trashed lifecycle status.

Automation: Generate Layer

An external MCP client can generate a sticker for an exact saved render without opening its viewer. This is local automation through the desktop app. The lifecycle is explicit: get_layer_draft → generate_layer with target: saved → get_layer_generation → land_generated_layer with target: saved → commit_layer_draft with target: saved. Generation produces a job artifact; landing adds it to the draft, and only the separate commit writes the Layers graph operation. Review and approve each change. A saved render currently owned by an open Layers session is not available to this path.

The saved workflow also provides cancel_layer_generation and rematte_layer; both default to execution_target: saved. Cancel targets the exact job and may report that cancellation was requested before the job reaches a terminal state. An external MCP client supplies the retained rgb_draft token for re-matte; Copilot resolves it privately for the approved exact job, without putting it in the conversation. The separate re-matte target field identifies a witnessed take to replace, not whether execution is saved or live. If a start response is lost, check the same request ID as the job ID; do not start a second job. A missing job after an app restart, or an interrupted land/commit, is an unknown outcome: inspect the saved draft and graph before trying again. rematte_layer starts a new job with a server-minted ID, so if its response is lost, inspect local job activity and the draft before making a new, explicit request.

In live Copilot, arm Layers on the intended render first and use Copilot's Automation mode. Approve generate_layer to start the job, then separately approve land_generated_layer after it is ready, and commit_layer_draft only when you want to bake the draft. The live actions are bound to the armed render and draft; Copilot will not arm or retarget them. Approved cancel_layer_generation or rematte_layer with execution_target: live controls only that exact armed session's job; a cancellation request is not proof that GPU work has stopped. Re-matte produces another artifact and, when given a witnessed recovery target, a later approved land replaces that take rather than adding one. Stop or switching away can refuse an action that has not been admitted. Once generation is admitted, Stop does not cancel that GPU job; it may finish as an artifact and will not land or commit by itself. If an admitted land, commit, or re-matte loses its acknowledgment, treat the result as unknown and inspect the exact job, Layers draft, or graph history before a new approved attempt. The Agent Mode & MCP guide covers connecting an external client.

Automation: Spot Clone

An external MCP client can work on an exact saved render without opening its viewer: preflight_spot_clone inspects readiness and mask/source fingerprints, then a separately approved clone_spot writes one node against the expected parent and HEAD. Keep the same node ID for reconciliation; if the reply is lost, inspect that exact ID with get_spot_clone before any new approval. A moved HEAD, changed mask, or changed source requires a fresh preflight. The client does not get to force a bake or silently choose another source or destination.

For live Copilot, first prepare Clone from… yourself in the Spot Selection card. Copilot's approved clone_spot can commit only that already prepared operation while its render, HEAD, source and destination poses, and mask fingerprint still match. It cannot arm Clone, pick a source or destination, redraw the mask, force, or retarget. Stop or switching away can prevent a clone that has not reached the write; it does not undo a bake already started. A completed clone is a graph node, so use graph history for Undo. If the result is uncertain, inspect the exact node receipt rather than retrying with a new ID.

Conversations

Every conversation is saved automatically and titled from its first exchange. The panel header has a history list to switch between them, plus a button for a new conversation. Delete any you don't want to keep.

Drag the panel's left edge to resize it. The width persists.

What Copilot knows about you

Each question is sent with a snapshot of your setup so answers are grounded in reality rather than guesswork:

• GPU model and VRAM
• Installed models and available templates
• License tier
• Your current render configuration and active pipeline
• Your 10 most recent renders, 5 most recent sequences, and 5 most recent 360 LoRA batches

None of this leaves your machine when you're using your own API key beyond what goes to your chosen provider.

Working offline

The product knowledge base is cached on disk for 7 days, so recent documentation remains available when 360hextile.com cannot be reached. This cache is documentation, not a local language model: Copilot can answer only while its selected model endpoint is reachable. The hosted assistant needs 360hextile.com; a BYOK model needs that provider. With no documentation cache, Copilot falls back to a minimal built-in reference after it reconnects to a model endpoint.

Examples

• "Why does my sky look muddy?"
• "Bump the steps and lower the guidance a bit"
• "Switch me to SD 3.5 Turbo for a fast test"
• "What pipeline should I use for landscapes?"
• "How do I fix pole distortion in my output?"
• "Open my most recent render"
• "Take me to the output settings"
• "What settings should I use for nighttime scenes?"
• "Align these layers to the horizon"
• "Set the selected logo to 70% opacity and move it behind the title"

Limitations

Copilot's answers come from documentation and your app state — it can be wrong, and it doesn't know about every edge case. Generate Layer and Spot Clone are limited to the explicitly approved saved or guarded live lifecycles above. Copilot cannot start an ordinary image render, prepare Clone from… for you, operate an unarmed or closed tool, or silently land or commit a result. For complex issues, see the Troubleshooting Guide or contact support.