Skip to content
For the complete documentation index, see llms.txt. This page is also available as Markdown.

Release 0.44.0

Previous Version: v0.43.0

Manage sharing and version checkpoints from the project toolbar, browse nodes in a clearer menu, and see concrete reasoning settings. Connected AI tools can now recover agent tasks after a disconnect, inspect status and budgets, retrieve private outputs, and package selected results into ZIP files. Available nodes and options depend on account and workspace access.

Manage project access from the Share panel

Click Share in the editor to manage workspace access and public links together. The panel shows the workspace and its member count, Private, Can view, and Can edit access choices, and Anyone with the link sharing with a copyable URL.

  • When changing a private project to workspace access, an inline confirmation explains how files are published. Files already introduced to the project, and files introduced while workspace access stays enabled, are permanently published to workspace Assets. Returning the project to Private does not withdraw those files.

  • Only the project creator can change workspace access. Editors can create or revoke public links; workspace viewers can inspect the settings and copy an existing link.

  • Link preparation shows loading, error, and Retry states. Project settings in the asset picker opens the same panel.

Workspace access and public links were already available; the new panel gives them a dedicated place in the editor.

Manage private, workspace, and public-link access from the dedicated Share panel. Workspace attribution is hidden.
The workspace-access confirmation explains which project files will remain available in workspace Assets.

Save a version checkpoint in one click

The current version now appears beside the project name. When you have changes and edit access, click Save as vN to save a checkpoint without a naming dialog. Project changes continue to autosave independently.

Click the project name to open project actions and, for editors, version history in one menu. Versions appear newest first with the saver, time, current-version indicator, and “based on” information where applicable. Open a version directly, or use its row menu to rename or delete it.

Opening a version affects everyone in the project. Existing restrictions during active work, and prompts to save or discard changes before opening another version, still apply.

The project toolbar keeps the saved version beside the project name.
Project actions and saved version history share one menu. Personal attribution is hidden.
Save a checkpoint, then browse version history.

Browse node groups side by side

The Add node browser presents groups as icon tiles with node counts, with sticky section headings inside each group. On larger pointer-based layouts, hovering or clicking a group opens its node list beside the top level. Compact and touch layouts use a single column with Back navigation.

Use the arrow keys and Enter to browse and add nodes. When you drag a connection, counts reflect compatible nodes and groups with no compatible nodes are dimmed.

Character tools are now together under Rigging & animation. Find rigging, motion generation, animation, retargeting, humanoid normalization, and FBX export in one group, subject to your existing node permissions.

Rigging, animation, retargeting, normalization, and FBX export appear together under Rigging & animation.

See concrete reasoning settings

Reasoning controls in LLM nodes and the project assistant now show the selected model’s actual configured level, replacing the ambiguous Default choice. Models that support disabling reasoning use Off consistently.

New nodes use model-specific defaults. Switching models keeps a supported selection or chooses the new model’s default. Saved graphs and exported API definitions migrate legacy selections to concrete levels, including old non-reasoning settings. API- and MCP-created workflows use the same model-specific defaults.

These controls update the behavior of existing models; they do not introduce new models.

Models that support disabling reasoning label the choice Off.
This model uses Minimal, Low, Medium, and High rather than an ambiguous Default option.
The project assistant uses the same concrete, model-specific reasoning choices.

Recover MCP agent tasks after a disconnect

The new start_project_agent tool returns an operation receipt while the agent works. Poll get_project_operation, or find earlier work with list_project_operations, to inspect progress, the final response, errors, and credit accounting without another model call.

Use a stable idempotency key for each request. Repeating the same prompt, requested budget, and key retrieves the same operation, helping you recover a lost start response without duplicate generation. Reusing that key with different inputs returns a conflict.

Accepted asynchronous work continues after the client disconnects. If a worker is lost, the operation is reported as interrupted; potentially paid work is not automatically replayed. The existing synchronous prompt_project_agent tool remains available.

Inspect MCP status and budgets without an agent turn

Connected tools can read project state and check a credit cap before starting agent work.

  • get_project_status reports node and selected-run IDs, running, scheduled, and error states, available warnings and errors, last-run cost and cache information, and the active asynchronous operation. Status reads stay available while the agent works.

  • preflight_project_agent reports the requested and effective credit cap and active operation without spending or reserving credits. It checks the cap; it is not a total-task quote or a guarantee that the workspace can fund the work.

  • Operation receipts distinguish settled spending, reservations, remaining credits, pending work, and whether accounting is complete. When an OAuth connection cap lowers a requested budget, a notice explains the reduction.

  • Project creation, list, and detail responses include an authenticated project_url that opens the editor directly. This is separate from a public share link and does not enable sharing. Compatible clients also receive clearer tool titles and action hints.

Retrieve private project outputs through MCP

Use get_project_outputs to retrieve generated files and authorized download links from any project you can view, including private projects. You no longer need to publish the project to workspace Assets or enable public sharing to retrieve its results.

Filter by node, run, or output role, such as base color, roughness, or normal. Paginated file descriptors include filenames, types, sizes, stored dimensions, node/run associations, and available generation provenance, including reuse of cached results.

Compatible MCP Apps hosts can show supported images in an inline results card with Download, Refresh links, and Next page controls. Other clients receive the typed manifest and URLs.

Download links last 15 minutes, open without an Authorization header, and recheck current credential and project access. Read the manifest again to refresh them.

A completed, organized image workflow whose generated result can be retrieved privately through MCP.
The MCP results card previews a generated image with Download and Refresh links controls.

Package selected MCP outputs into a ZIP

Use create_project_output_bundle to package selected existing output files, then get_project_output_bundle to follow progress and download the result. Choose the ZIP filename and each file’s relative path inside the archive. A manifest.json connects archive entries to their output metadata and provenance.

A bundle supports up to 100 files, with 512 MiB limits on both the combined source size and the completed ZIP, subject to the existing project ZIP quota. Packaging uses existing file bytes without invoking a model or regenerating results.

Creating a bundle requires edit access because the ZIP is saved in the project. In a project already published to the workspace, the ZIP is published like other project files.

The completed ZIP bundle is available to download from the project.

Change a connected tool’s credit cap without reconnecting

On Workspace → MCP, the connection owner or a workspace owner/admin can edit Credits per agent call and choose Save limit. Choose 250, 500, 1,000, 2,500, or 5,000 credits.

New calls use the saved cap. Already accepted calls keep their original budget, including when recovered with their original idempotency key. The connection and its existing download links stay valid, without reconnecting or rotating credentials. This control applies to active OAuth connections.

Change Credits per agent call and save the limit on an existing connected tool. Personal attribution is blurred.

More consistent pages and forms

Cards, form fields, table rows, badges, and separators now share a more consistent visual hierarchy across projects and workspace settings. Flatter rows and clearer input backgrounds reduce layers of nested panels.

The update also covers API keys, plans and credits, usage, transactions, hotkeys, pricing, plugins, exported APIs, and invitation acceptance.

Flatter project cards and clearer search controls in the refreshed Projects page.
Consistent fields, rows, and separators in workspace settings.

Reuse nodes from Recent and Favorites

The Add node menu now includes Recent and Favorites. Recent keeps the last six distinct node types added from this menu, newest first. Bookmark a node in a browsing list to save it without adding a node to the graph. Favorites keeps up to 24 node types, with newly saved entries first; remove a bookmark to take a node out.

Both lists are saved to your account and follow you across projects and browsers. They show nodes available to you, and while dragging a connection, their lists and counts show compatible nodes. Empty lists explain how to get started.

Keyboard users can press Space on a highlighted browsing node to toggle its bookmark and Enter to add it.

Bookmark a node in the Add node browser, then find it in Favorites beside an organized workflow.

Editor panels, menus, dialogs, and notifications now use more consistent backgrounds and separators. Model choices have clearer group headings and selection styling, with a dedicated Search models field.

The project assistant model picker has a clearer Search models field, grouped choices, and a visible selected model.

Last updated