Manual · Workspace

Workspace & asset library

The workspace top bar, four-pane layout and new-asset entry points; the AI chat panel, skills and the plugin marketplace; and browsing, filtering, context menus and shortcuts in the asset library.

5. Workspace

Opening a project takes you to the workspace (named “Workspace” in the app navigation). The interface is a dockable multi-panel layout.

5.1 Top bar

Unsaved changes show * next to the title; press Ctrl+S to save.
↶ / ↷: undo / redo (also available as shortcuts).
One-Click Workflow: generate a reusable host asset from a preset or AI planning (see §7).
Episode Pipeline: open the short-drama Agent pipeline window to work through 9-grid, 4-grid, director review and video generation (see §7.5).
Task list and Run log: background queue and model call logs.
Layout: switch the default layout, save the current layout, export / import a layout .json, delete custom layouts.

Tip: drag a tab to dock it in another area; right-click a tab to float it, detach it into a new window, or close related tabs. Key panels can't be closed while nodes are running.

5.2 The four panels

Panel	Purpose
Tools	Narrow left column with shortcuts for creating asset drafts.
AI chat	Opened via the “◈” button at the bottom of the left column; an in-app AI assistant that can call MCP tools in conversation (see §5.5).
Workspace	The central editor; the empty state shows the recommended flow and recent assets, and opened assets become tabs.
Assets	Browse and manage the project asset library.
Parameters	Inspector panel for the current selection.

5.3 Left toolbar: create assets

Click an icon to create a draft asset, then press Ctrl+S after editing to write it to disk. You can create:

Free canvas · host asset · script · world elements · shot · director stage · image · video · audio

Recommended flow in an empty workspace: script → shot → node generation; you can also create a script, shot or director stage directly, or double-click a recent asset to open it.

5.4 Right rail: assets / parameters

The far right of the workspace is a Rider-style vertical rail; the Assets and Parameters panels can be expanded or collapsed independently (a collapsed panel takes no width from the center layout).

Click a rail button to toggle the matching side panel.
You can drag “Assets” above or below “Parameters”; drop previews only appear in the top / bottom halves (half-screen left / right previews are disabled to avoid leftover grey frames).
The layout menu saves the expanded geometry, while collapse preferences are remembered separately — a collapsed state is never written as a zero-width grey hole.

5.5 AI chat panel

The ◈ AI chat button at the bottom of the left column opens the in-app AI assistant. It runs the DeepSeek Harness (dsh) agent runtime and can call the app’s built-in MCP tool server exactly like an external agent — in other words, drive generation and project operations right from the chat window.

@-reference assets: type @ to open an asset picker and inline project images / videos / audio / 3D models into your instruction (the composer shows thumbnail previews). dsh parses the @path references directly.
Call generation tools: ask “generate a title image from @bg.png into the project” — the assistant calls MCP tools like generate_image, and tool progress is shown as live cards (running / done / failed).
Model picker: at the bottom of the panel you can pick any configured text model with a key (not limited to DeepSeek official); the choice is remembered.
Multi-session: the toolbar above the composer creates / switches / deletes sessions; message history persists locally.
Cancel & status: you can stop a running turn; dsh status lines and tool activity stream back into the panel in real time.
Approval when the agent needs to go past the sandbox: a consent card appears in the conversation — “This step needs your approval” — naming the tool and the reason, and offering only Allow once / Reject. A grant applies to that single call: the app deliberately offers no “always allow”, so one click can never quietly widen what the agent may run later. When no one is available to answer (for example fallback one-shot execution paths) the request fails closed: it is rejected, never silently allowed.

Prerequisites: AI chat bundles both the Node.js runtime and the dsh runtime — ready to use out of the box, no system Node.js required (only exotic environments fall back to on-the-fly downloads). You do need at least one enabled text model with a key; without an open project only pure conversation is available.

5.6 Skills

The chat agent ships with a set of creation skills (derived from the app’s built-in storyboard / director capabilities, injected automatically on each conversation); you can also add your own skills in the Skills tab of the plugin marketplace to give the assistant a tailored way of working.

Built-in skills: storyboard / animation (9-grid storyboard table, beat breakdown table, motion prompt table, 4-grid motion storyboard table, …), director review (storyboard / beat / motion prompts), and system creation (script, image-from-prompt, image / video generation, voice, emotion, lighting, multi-angle, upscale, redraw, erase, HD enhance, UI image, UI split, world extraction, beat split, beat unit generation, plan, portrait texture, prompt optimization). These are managed automatically — please don’t edit them.
Custom skills: in the Marketplace → Skills tab click “Open folder” and drop in a .md file in dsh SKILL.md format (frontmatter name / description + Markdown body; kebab-case filename such as my-skill.md). It takes effect on the next conversation. “Create sample template” in the same tab gives you a ready-to-use reference.
How skills work: the skill catalog is injected into the conversation context as <available_skills>; when the agent judges a task matches, it loads the skill file on demand as instructions — the skill defines the process and conventions (how to do it), while MCP tools do the actual work (generate images, run workflows).
Node-graph nodes use the same skill system to define their behavior; API run logs annotate the skill used by each node.
Skill bundles: installing a workflow from the plugin marketplace may bring a skill bundle with it. It is listed in the marketplace’s Skills tab with its source marked as a skill bundle, and it belongs to the workflow that shipped it — uninstalling that workflow deletes it, so it is not something you manage on its own.

Skills vs tools: skills are the “job manual”, tools are the “hands”. Say “use the storyboard skill to split this script into a 9-grid storyboard table” — the skill provides the professional workflow (beat → framing → motion → prompt) and tools execute the generation.

5.7 Plugin marketplace

“Marketplace” sits immediately to the left of “Settings” in the top bar and opens a separate window where MCP servers, skills and workflows are all installed from. It is an app-level entry — no project needs to be open — and the window is a singleton: clicking again only focuses the window that is already open.

Its tabs are All / MCP / Skills / Workflows. Every time the window opens it force-refreshes the workflow catalog, so a newly published workflow shows up without waiting for a cache TTL; if that refresh fails it keeps the list it last read and shows a soft hint rather than an empty market. The catalog has two sources — the official primary (GitHub raw) and a Gitee mirror — tried in order with automatic fallback. The top of the window says how many entries were read (when it fell back it says “Primary source unreachable — switched to the mirror”), and offline it says the cached catalog is being shown.

Want to publish your own workflow or skill bundle? See the plugin marketplace developer docs.

Install, update and uninstall

Workflow cards carry Install / Update / Reinstall / Uninstall and Details right on the card; installed entries show their status, and “Update available” appears when a newer version exists.
Uninstall goes through the app’s own confirmation dialog; once confirmed, the locally installed files are removed. If the workflow shipped a skill bundle, the uninstall removes that skill too.
The market only installs onto this machine: use the workflow in AI chat (below).

Entries that include a skill, and script consent

A workflow may ship a skill bundle: the card shows an “Includes skill” badge, or “Includes skill + scripts” when the bundle also carries scripts — visible before you install.
Scripts in a skill bundle are code the AI agent can run on this machine, so they are written only with your explicit consent: installing asks first, listing every script path and stating plainly what these scripts are. Confirming writes them; cancelling installs the instructions and references only — no scripts on disk, and the workflow still works.
That consent is asked every time: the app never remembers an “always allow”.
Installed skill bundles show up in the marketplace’s Skills tab, with their source marked as a skill bundle; the AI chat agent can load them directly. A bundle belongs to the workflow that shipped it — uninstalling that workflow deletes it too — so it is not something you manage on its own.

The MCP tab

The built-in MCP tool server (aiartengine) and the Blender toolset are inspected here — see §4.5 and the MCP setup guide.
You can also add third-party MCP servers: for HTTP (remote) enter the service URL and headers; for stdio (local command) enter the command, arguments and environment variables. “Test connection” before saving shows how many tools the server exposes. stdio runs a third-party program on your machine, so only add services from sources you trust.

Using an installed workflow in AI chat

Two entry points: the “Workflows” button in the chat composer toolbar, or by typing /workflow.
Selecting a workflow inserts a reference block (a pill with icon + title) into the input box — it deliberately does not paste the title / id / summary as a paragraph. Hovering shows the full reference including the id.
On send, the agent uses the installed workflow by id (workflow_use_installed), reproducing that exact workflow rather than re-planning something similar.

Two gates: script consent at install time decides whether code is written to disk; when the agent actually needs to go past the sandbox there is a second gate — the “This step needs your approval” card in the conversation, offering only “Allow once / Reject”, with no “always allow” and with unanswered requests failing closed (see §5.5).

6. Asset library

Browse and filter

Refresh next to the “Assets” title; search and type filters are supported.
The slider at the bottom switches between list view and icon view.
A breadcrumb shows the path, e.g. Assets > …
Marquee multi-select is supported; you can drag images / videos / audio / models / .aipackage in from your system.
Multi-select files / folders dragged to another directory: left-side directory rows and right-side asset cards are both draggable; drag a batch of files or multiple folders onto a target directory at once (everything goes through the folder_move IPC + the same on-disk rules: physical move + descendant asset.relativePath rewrite + cycle detection + duplicate-name auto-append " 2" / " 3"). Dragging within the same parent only rewrites .folder.json and doesn't move any data on disk. The asset library root cannot be dragged out of.

Right-click empty space or a folder

Create assets of any type, create a folder.
Import asset package / Export asset package.
Folders additionally offer: reimport, rename, delete folder (contents move up), delete folder with contents.

Right-click an asset

Open editor · reveal in file manager · copy original file.
Reimport · export asset package · rename · find references · delete.

Shortcuts with the asset library focused

Delete / Backspace: delete the selection.
Ctrl+A: select all.
Ctrl+C: copy-original-file related action.