Learn Mixar

Your first project

Start with a small object so you can learn the cycle: describe, inspect, refine, and save.

Set up

Download Mixar for macOS, Windows, or Ubuntu Linux and install it. You can start in guest mode, or sign in to use your account. Start a project and save it before making substantial changes.

Build your first object

1. Open the floating Agent from the Mixie pill. Choose your own AI provider, or sign in and choose a subscription to use Mixar credits.
2. Send: “Create a small round side table with three wooden legs. Keep each part editable and place it on the ground.”
3. Answer any questions in the conversation and follow the task progress.
4. Orbit around the result. Check that the legs meet the tabletop and that the table sits on the ground.
5. Refine one thing at a time: “Make the tabletop thinner. Keep its diameter and the legs unchanged.”
6. Save the project again.

What you should learn

A useful follow-up names both the change and what to preserve. The conversation guides the work, while the saved project holds the scene. A completed reply does not replace saving.

Start without signing in

Guest mode lets you open Mixar, build and save projects, and try its editing tools before creating an account. To chat with the Agent, connect your own AI provider. You can also connect an external AI app to use free scene tools.

Sign in when you want to use Mixar-hosted Agent models or paid generation. Your provider or external AI app may charge for its own usage even when no Mixar credits are used. Guest mode still needs an internet connection for connected features.

When you sign in, your guest conversation carries over. Your guest provider setup carries over if your account does not already have one. Save your project to keep your scene work.

Zen and Engine

Use the Zen / Engine switch in the top bar to choose how much of the editor you see.

• Zen keeps the viewport, transform tools, floating Agent, and Moodboard drawer close at hand.
• Engine exposes the full Blender-style workspace for mesh editing, UVs, materials, animation, and detailed scene management.
• Cinema Mode opens the camera-directing workspace.

These are different views of the same project. Switching modes does not create a separate copy of your scene. Use the viewport shading controls to inspect geometry, materials, or rendered lighting.

Find your way around

Orbit to check a model from more than one angle; pan and zoom to inspect details. Select an object before asking for a change to “this object,” or use its name to remove ambiguity.

The floating island brings Agent, 3D, Image, Video, Splats, Add-on, Library, and Queue together. Open the Moodboard tab at the right edge of Zen to work with references beside your scene. Drag the drawer grip to change its width.

If you are new to the interface, use Help → Welcome to revisit the guided introduction.

Write a useful request

Use the Agent when you want to build or change something in the scene, or coordinate several creation steps.

Give the request a clear shape

Include these four pieces:

• Target: the selected object, a named object, or a new object to create.
• Change: what should be different when the task is finished.
• Constraints: dimensions, placement, materials, or parts that must stay unchanged.
• Result: an editable model, a generated asset, a material, an image, or an animation.

For example: “Make the selected chair seat 10 cm wider. Preserve the legs and wood material, and keep the seat editable.”

Add context and send

Use Attach to add reference images, a 3D file, individual files, or a project folder. Plain-text notes and briefs can guide a request; a video can provide motion reference. Images appear beside the composer. A 3D attachment imports into the active scene; removing its attachment does not remove the imported objects. Explain each reference’s role: “match the shape” and “use these colors” ask for different things. Select Send, or press Enter; use Shift + Enter for a new line.

The Agent can work on modeling, UVs, materials, lighting, cameras, animation, scene organization, and generation workflows. For PDFs or Word documents, paste the relevant text or attach page images: attaching the file does not make its contents readable automatically.

Inspect and refine

Compare the result with the constraints in your request. Check the whole object from several angles, then give a focused correction. “The back legs are too close together” is easier to act on than “fix the chair.”

Use History to revisit a conversation and New chat to start a separate discussion. Starting a new chat keeps the existing scene.

Refine and reuse a prompt

Write a draft, then use Refine beside Send to improve its wording. Review the result before sending; Revert brings back the previous draft if you prefer it. Refining a draft does not send it to the Agent.

Use Up / Down in the composer to recall earlier prompts and return toward your current draft. Review recalled text and attachments before sending. Longer briefs fit in the composer, but a clear target and a few specific constraints are usually easier to refine.

Use a video as motion reference

Attach a video or include it in a project folder, then describe the movement to study. For example: “Compare the character’s walk with this clip. Focus on stride length and foot contact.”

The Agent can inspect selected frames from a reference video or a rendered clip. Mention useful timestamps and use a model that supports images. Frame inspection helps compare poses and motion; it does not mean the Agent has watched every frame or listened to the audio.

Find references on the web

Ask the Agent to look up references, consult a public page, or bring a public file into your project. For example: “Find references for a mid-century desk lamp, show the sources, and use them to guide the shape.”

Web search and page reading depend on the selected Agent model. If they are unavailable, choose a supported model or provide the reference directly. The Agent can also download public images, models, textures, or other files up to 50 MB from a direct link. Pages that require a private account are not accessible this way.

Review the sources and check that downloaded assets are suitable for your project. Search usage can contribute to Agent charges; with your own provider, that provider handles its usage charges.

Work in separate scenes

Open Scene at the left of the Zen toolbar, or press Ctrl + backtick. The same drawer is available in Engine.

1. Choose + New scene to start a separate scene and Agent conversation in the same project.
2. Click a scene card to switch. Read its status to see whether it is working or waiting for you.
3. Use the card’s drag grip to reorder scenes. The Agent header’s scene-tabs control also lets you switch scenes or send selected objects to another scene.

Each scene has its own conversation and can continue work while you inspect another tab. Generation and Detect Views results return to the scene that started them, even if you switch tabs. Check the active scene before submitting instructions, and save the project to preserve your scenes.

Undo within one scene

Undo, Redo, and Undo History follow the current scene tab, so stepping back in one scene preserves work in the others. Switch to the scene you want to change before using them.

Use Edit → Undo Whole Document when you deliberately want to step through changes across the project. If a change involves data shared by scenes, Mixar may ask you to use whole-document undo instead of changing only one tab.

Undo does not rewind the Agent’s conversation. Use Checkpoints when you need to recover an earlier Agent turn and its document state.

Keep project instructions in Rules

Open Rules, the document icon beside Checkpoints in the Agent header. Add reusable instructions such as “Use meters and keep object names descriptive.”

Choose This project for instructions tied to the current project, or All projects for a preference across Mixar files on this computer. Review the selected scope, then add or save the rule. You can edit, disable, or remove rules later. A one-time request belongs in the conversation.

Give the Agent folder context

Choose Attach → Project folder and select the folder the Agent should use. Its folder chip appears with the other references. Describe the relevant files and what you want done, such as “Use the reference images in this folder to match the furniture colors.”

Attach only the folder needed for the task. A context folder lets the Agent inspect relevant local files; it is different from connecting a folder in Library. Remove its attachment when you no longer want it included in your request.

Questions and planning

For a complex change, ask the Agent to explain its approach before making edits. State your approval preference in the request, such as “Explain the proposed changes first and wait for my approval.”

Auto in the composer lets the Agent decide open choices and summarize its decisions. Leave Auto off when you want to answer those choices yourself. An explicit request to wait for approval still matters.

When Mixar presents a question, select the appropriate answer or write a correction. For a sequence of questions, complete the set before the Agent continues.

Progress and follow-up instructions

A larger request may be split into several tasks. The task cards help you follow the work while the Agent coordinates the result. A task finishing its build is not the same as the whole request passing review.

Some work continues after an intermediate reply. Keep Mixar open and connected while it finishes; generated assets may take longer than scene edits. You can send a clarification during an open run, such as “Keep the lamp brass, but make the shade cream.”

Read the final summary and inspect the scene. If Mixar offers Retry failed tasks or Review unfinished checks, use that action to continue the unfinished work.

Go back with checkpoints

Use a checkpoint when you want to return to the document as it was before an earlier Agent turn.

Before restoring

Wait until the session is idle and no work remains open. Decide which turn introduced the change you want to remove. Save a separate project copy if you want to keep the current version independently.

Restore an earlier turn

1. Open Checkpoints, beside History and New chat.
2. Choose the point before the change you want to discard.
3. Wait for the restore to finish and read any recovery notice.
4. Check the scene, then restate your next request in full.

Mixar snapshots the document before a fresh turn, retains the latest 20 checkpoints per session, and creates a safety copy before restoring. Identical document states can share a snapshot.

Restoring also saves the restored document to the project’s file. Save a separate copy before restoring if you need to preserve your current version. Checkpoints are recovery points, not a substitute for your own project backups.

What changes after a restore

When the matching conversation checkpoint is available, the Agent’s conversation also rewinds. It no longer remembers discarded requests or results, so “try that differently” may have no meaning. Describe the new task again.

A scene-only recovery notice means the conversation did not rewind too. Images linked from disk still use the pixels currently in those files; a scene checkpoint does not rewind external image files.

Stop, reconnect, and recover

Use Stop to interrupt the current work. Stopping does not undo changes already made to the scene. Review the result, then continue with a corrected request or restore an available checkpoint once the session is idle.

A temporary connection loss is not proof that the work was cancelled. Give Mixar time to reconnect and recover the run before submitting the same request again. If the app cannot recover, save your project and follow the error message.

For repeated failures, report a bug with the request, app version, and steps that led to the issue.

Point at a part of the scene

Use a viewport mark when words alone would make the target unclear, such as a small handle or one part of a larger object.

Mark the target

1. Frame the object in an ordinary viewport view. Exit camera view first.
2. Choose Sketch in the Agent. The viewport freezes and the Agent minimizes while you draw.
3. Circle the part, point at an object, or mark an empty spot on the ground.
4. Choose Done or press Escape over the viewport to return to the composer and inspect the sketch preview. Add an instruction, such as “Make the circled handle thicker, keeping the rest unchanged.”
5. Send, then inspect the changed area from more than one angle.

For placement, mark the ground and ask “Place a small plant here.” Press Escape over the viewport to finish drawing and keep the queued sketch. Ctrl / Cmd + Z undoes ink; Backspace edits your prompt text. Type over the frozen viewport to add instructions, or reopen the pill. Click a sketch preview to inspect it at a larger size; use its × to discard that draft sketch before sending.

Pointing or sketching?

A pointing mark says where to work; your words say what to do. If your drawing represents a new object or arrangement to build, use the Sketch workflow.

Build from a sketch

Draw the arrangement you want over the frozen viewport. Choose Draw to build when the ink represents a new object, or Point to edit when circles and arrows locate a change. Auto detect lets Mixar infer the intent. Tab switches the interpretation while the pointer is over the frozen viewport; check the hint before sending. Finish with Done and inspect the preview before sending.

Add words for details a sketch cannot communicate, such as scale, material, or which elements should be separate. For example: “Build this desk layout. The two side rectangles are drawers; leave an open space in the middle.”

Sketches guide the layout and intent. Inspect the result from several angles and refine it with follow-up requests.

Handwriting and voice

Click the Handwriting signature icon in the Agent header to open the writing canvas and turn handwriting into composer text. This is different from marking the viewport: handwriting becomes the words of your request. Review the recognized text before sending.

Where microphone capture is supported, use Voice when you prefer to dictate. You can also hold left Option on macOS or left Alt on Windows in a text field, then release to insert the transcript. Check the transcription for object names, numbers, and units. A spoken or handwritten request benefits from the same clear constraints as a typed one.

Add and organize references

Use the Moodboard to collect the images, videos, and notes that explain your project’s direction.

Set up a useful board

1. Open the Moodboard drawer in Zen, or the full Moodboard editor in Engine.
2. Drop images or videos onto the canvas, or use the folder control to add media.
3. Arrange related references together. Add text notes explaining which image defines shape, color, material, or composition.
4. Select a still image to attach it to the Agent automatically. Selecting a frame includes its still images, up to the reference limit. Deselect an image or remove its attachment to stop including it. For a generation node, check its selected or connected input separately.

Pan and zoom to keep your working group visible. Drag the drawer grip to give the board more space.

Check the input before generating

A board can contain many references, but a generation step may use only its selected or connected input. Confirm that the intended image is selected rather than assuming the whole board will be used.

Moodboard annotations organize ideas. Use viewport Sketch to identify a specific part of a 3D object.

Select regions with masks

Select a still image, then open Mask Tools on the Moodboard toolbar. Choose Box Mask, Multi-Lasso Mask, or Magic Select for the region you need. These tools are also available in the full Moodboard editor.

Masks stay with their source image and can define named components. Inspect the selected area before using a compatible generation action; the available actions and models determine how a mask can be used. Press Escape to leave the tool. Viewport Sketch marks identify 3D scene targets; Moodboard masks select image regions.

Connect steps on the canvas

Connect steps when one result will become the input to another, such as reference image → 3D model → retopology.

Build the sequence

1. Start from the source media card and use its + handle to choose a compatible next action.
2. Check the connection created for the new action. To connect existing cards, drag the source’s + handle onto a compatible input or action card.
3. Review the input and model settings, then run the action.
4. Wait for the result and inspect it before starting the dependent step.

Available actions depend on the source media and enabled models. You can branch from one image to compare several treatments without replacing the original reference.

Know what a connection means

The line supplies an input; it does not mean a job has finished. Check Queue and inspect the actual result. If the source image is unsuitable, fix it before spending time on downstream geometry or textures.

Generate and edit images

1. Open Image in the floating island, or Image Gen in the Moodboard tools.
2. Choose an available model and describe the image.
3. Add a reference when the chosen model supports one. Specify what to keep and what to change.
4. Set the available aspect ratio and image count, then choose Generate.
5. Find the result on the Moodboard and inspect it before using it downstream.

For a 3D reference, use a single subject, a simple background, and the entire object in frame. Avoid text and surrounding props unless they belong in the model.

Model choices and settings come from the live catalog. A screenshot illustrates the workflow; your available models may differ.

AI Render: explore a look

Build a rough scene with visible geometry, frame the view, and use the AI Render workflow to describe a visual treatment. You can also capture the viewport as a reference with Capture Viewport in the Image tab.

For example: “A warm evening interior, soft window light, pale oak and linen.”

The result is an image for look development. It does not automatically change the scene’s materials, lights, or camera. Ask the Agent to make those scene changes when you want an editable 3D result.

Generate video

Open the Video tab. Choose an enabled model, provide its required prompt and source media, and review the available settings before generating.

Review duration, resolution, aspect ratio, audio, and reference-video options where the selected model exposes them. When using multiple Moodboard references, check their displayed order and labels, such as @Image1 or @Video1, and name each reference’s role in your prompt.

Video inputs vary by model. Some workflows use an image or first frame; use the fields actually shown for the selected model rather than assuming every model accepts the same inputs.

Generated video is media on the Moodboard. To create editable camera motion inside your scene, use Cinema Mode instead.

Choose the right starting point

Use 3D for generated meshes, Image or Video for media, and Splats for a splat environment. You can also ask the Agent to coordinate a generation workflow.

For precise editable construction, say so: “Model this as separate editable parts.” For a generated asset, provide a clean reference and state your geometry and texture priorities.

Read the selected model’s settings before submitting. Face limits, texture options, quality levels, and supported inputs vary by model. Check the app’s current credit information; this guide does not prescribe fixed prices or generation times.

Image to 3D and multi-view

Use this workflow to reconstruct a mesh from a visual reference. It is a useful starting point when appearance matters more than specifying every modeling operation.

Prepare the reference

Choose a clear image with one complete subject, a simple background, and little occlusion. Hidden surfaces still need to be inferred, so a good front view does not guarantee a correct back or underside.

Generate the mesh

1. Open 3D → Image to 3D.
2. Upload the reference, or select the intended Moodboard image where supported.
3. Choose a model. Review its texture, PBR, quality, and geometry settings.
4. Select Generate and follow the job in Queue.
5. Find the imported mesh, select it, and frame it in the viewport.

Check before refining

Inspect the silhouette, back, underside, thin parts, scale, and materials. Correct large shape problems before investing in retopology or detailed texturing. Keep the original result when comparing alternatives.

Using multiple views

A turnaround sheet shows one subject from several angles. Use a model and workflow that support multiple views, with consistent front, side, and back images. Treating a multi-panel sheet as an ordinary single-view image can produce a mesh that combines the panels instead of reconstructing the subject.

Segments and character parts

There are two different starting points:

• An image: use a segmentation or character-parts workflow to identify subjects or generate an asset split into parts.
• An existing mesh: select the mesh and use Mesh Segment to create separate part objects.

Inspect the returned parts and names. Mesh segmentation can hide the original while importing a parts collection; it is not simply renaming the original mesh. Keep a saved copy before replacing a working asset.

Retopology

Retopology creates a new mesh with a different polygon layout or density. Use it when an existing generated, dense, or scanned model needs a more practical mesh.

Prepare and run

1. Save the project and select the intended mesh.
2. Open Retopology in the generation tools, or ask the Agent to retopologize the named object.
3. Choose an enabled engine and a target density suitable for your next step.
4. Generate and wait for the new mesh to arrive.

Compare the result

Check the silhouette, holes, thin parts, small details, and materials against the original. If the asset will deform, inspect the edge flow around joints and test it in a pose. A lower polygon count alone does not make a mesh animation-ready.

The service returns a new mesh. Keep the original until the replacement passes your checks. A face target is approximate, not an exact-count guarantee.

Tripo retopology requires at least 500 faces. Use an in-scene simplification workflow for a coarser result below that threshold.

UV unwrapping and auto-rigging

For editable UV work, ask the Agent to mark seams, unwrap, inspect overlaps, and pack islands at a stated texel density or resolution. Inspect the checker pattern before texturing.

AI UV Unwrap is a generation workflow that returns a new mesh with UVs. Preserve the original and check the imported result.

For a character or creature, use Auto Rig on the intended mesh. Tripo Auto Rig 3.0 is the default Tripo option and supports humanoids and creatures. Choose Skeleton → Tripo native if you want to use its animation presets. If the character is already split into parts, include the parts as one character instead of generating a separate skeleton for each part. Test the resulting pose and deformation; automatic weights may need adjustment.

Add preset character animations

Give a compatible Tripo-rigged character a walk, run, idle, dance, or another available motion. Ask the Agent, for example: “Add a walk and an idle animation to this character, then preview the walk.” The available presets depend on its skeleton.

On the Moodboard, choose Animate from a compatible rig’s next-step menu, select the offered animations, and run the node. The clips are added to the existing rig so you can inspect them on your character.

Imported rigs from other tools, older humanoid rigs, and rigs using other bone naming may not support these presets. If Animate is unavailable, ask the Agent to check compatibility before re-rigging. You can still create custom animation with the Agent or Engine’s animation tools.

Browse animations in Zen

Select a character or animated object and open Animations from Zen’s object controls.

• Choose Selected object for that object’s clips, or All in file to browse other animations in the project.
• Search by name, then play a compatible clip to preview it. Clips that do not fit the selected object explain why they cannot play.
• Use the playback controls to pause, stop, change looping, or inspect a frame. Add a compatible clip to the object when you want to keep it there.
• Choose Hide Rig to hide the skeleton while the character keeps moving; Show Rig brings it back.

Check the motion from several angles before rendering or exporting.

Scenes and Gaussian Splats

Use the Agent when you want an editable scene assembled from objects. Describe the arrangement, scale, camera, and what must remain individually editable.

Use Splats when you want a generated environment represented as a splat. Choose image or text input as supported by the current model, set the available detail level, and generate. A selected Moodboard reference must be the image you intend to use.

A splat is not an ordinary polygon mesh. Do not assume mesh editing, topology tools, or mesh export formats work on it. Treat splats as a separate environment workflow and check the supported output options for your intended destination.

Follow jobs in Queue

Open Queue to follow submitted generation jobs and their status. Image, mesh, texture, and video jobs may finish at different times.

Wait for completion and inspect the delivered asset before starting a dependent task. If a job fails, read its error before submitting another copy. Check the selected input, supported model settings, connection, and available credits.

A queued request is not a completed result. Keep the project open while you are waiting for assets to arrive. Switching scene tabs does not change the destination of an already-submitted job.

Open the Library

Choose Library in the floating island to browse your generations and connected asset libraries. Use the available source filters and search controls to find the asset you need.

Choose the correct destination when adding an asset: an image reference belongs on the Moodboard; a 3D asset belongs in the scene. Inspect imported scale, orientation, and materials.

Connect and search your assets

In Library, choose My Libraries, then + Add Library… and select your asset-library folder. Connected folders can show images and videos as well as Blender assets. Plain image and video files can be added to the Moodboard. A loose mesh file does not become a marked Blender asset just by being placed in the folder; import it through the appropriate file workflow.

Wait for the asset list and previews to load. Select the intended library and use the available filters to narrow the results. Drag a 3D asset into the viewport to place it. For generated images or videos, use Select on Board to find the media on the Moodboard.

If the library is empty, confirm that you chose the right folder and that it contains supported media or marked Blender assets. Use Ctrl / Cmd / Shift-click for multiple selections, then apply the available batch action. Save Preferences if automatic preference saving is disabled and you want the library connection to persist.

Reuse assets with the Agent

Tell the Agent when you want to reuse a specific object already in your scene, and name it when possible. For a library asset, bring the chosen item into the project and then ask for placement or edits.

For example: “Duplicate this chair around the table, with enough space between each chair.” This gives you control over the source asset and lets the Agent handle the arrangement.

Your generations

Use AI generations to revisit available generated assets. Check which result you are importing, especially when a prompt produced several variations.

Save useful results in your project and keep any external files needed by your workflow. The scene you saved is the reference for your final assembly.

PBR channels

A material can combine Base Color, Roughness, Metallic, and Normal information, with additional channels where configured. Change the channel that describes the effect: roughness controls surface response, while normal detail changes the appearance of small surface features.

Select the object and its intended material before editing. Inspect it in Material Preview or Rendered shading; a flat solid view does not show the final material.

Build an editable layer stack

A layer stack separates a material into changes you can adjust independently. Start simple, then add detail only when the base surface works.

Build the material

1. Select the mesh and intended material in the Texturing workspace.
2. Initialize a Mixar Paint material if needed.
3. Add a fill layer for the base surface. Set its color, roughness, and metallic response as appropriate.
4. Add separate paint, image, or procedural layers for variation and detail.
5. Name each layer by its purpose. Adjust opacity and blending, and enable only the channels it needs.
6. Add a mask when a layer should affect only part of the surface.

Understand the layers

• Fill: establishes a surface across the material, or across the area allowed by its mask.
• Paint or image: adds image-based detail and brush work.
• Procedural: adds patterns or variation you can adjust through parameters.
• Group: organizes related layers so the stack stays understandable.

Check the result

Toggle layers individually to see what each contributes. Inspect the surface in Material Preview or Rendered shading under useful lighting. If wear changes the whole base color unexpectedly, check that layer’s enabled channels and mask.

For Agent assistance, ask for the editable structure explicitly: “Create an editable brushed-steel material with a separate subtle scratch layer.”

Control where a layer appears

Add a mask to restrict a layer to part of the surface. Paint a mask for local control or use a procedural mask for a repeatable pattern. Inspect the mask and the shaded material together.

If an effect covers too much, adjust its mask instead of changing the base material. Confirm which layer and mask are active before painting.

Paint and refine

Select the mesh, material, layer, and paint image you want to edit. Check the UVs, then choose a brush and paint. Work at a resolution suitable for the asset’s final use.

Keep broad color changes, dirt, wear, and fine detail separate so they can be adjusted independently. Save modified image assets as well as the project.

For generated brush textures, first enter Texture Paint with an active image-paint brush. A brush-generation job needs that paint context.

Generate textures for a mesh

Use Texture Gen for a generated textured result from an existing mesh. Describe the surface, such as “brushed steel with fine horizontal grain”, and provide a material reference when supported.

The whole-mesh texture generation workflow returns a new textured mesh. It does not create the same editable stack as layer painting. Keep the original and compare the result before replacing it.

If you want a material you can keep editing layer by layer, use the layer-paint workflow and state that requirement in your request.

Bake mesh information

Bake the mesh maps required by your material workflow, such as ambient occlusion or curvature where available. Confirm the source mesh, target mesh, UVs, and resolution before baking.

Check the baked images for seams, missing islands, projection errors, and incorrect scale. Rebake when geometry or UV changes invalidate the previous maps.

Bake and export materials

When the editable material is ready, bake its channels into the images required by your destination. Check the output resolution, channel selection, normal convention, and file locations.

Inspect the baked result on the mesh before export. Asset exports include baked Mixar materials; keep the editable project so you can return to its layers later. Reopen the exported asset in the destination application to confirm that textures are found and interpreted correctly.

Compose a shot

A shot combines a camera, framing, and the scene it sees. Establish a useful still composition before adding motion.

Frame the first shot

1. Choose Cinema Mode in the top bar.
2. Create or select the camera for the shot.
3. Set the aspect ratio and lens for the composition you want.
4. Position the camera, then check the subject’s size in frame, the background, and the edges of the image.

From the same camera position, a shorter focal length shows a wider field of view and a longer focal length shows a narrower one. Moving the camera changes perspective, so check the composition again after repositioning it.

Before recording motion

Confirm the active shot and camera. Leave room for the subject throughout the move, and check for objects that could block the camera’s path. Then continue to camera keyframes.

Camera keyframes and output

Camera keyframes record positions along a timeline. The preview between them reveals the motion, so review the transition as carefully as the endpoints.

Make and review a move

1. Select the shot and move to the starting frame.
2. Frame the camera and capture its first keyframe.
3. Move to a later frame, reposition the camera, and capture the next keyframe.
4. Scrub between the keys and play the move. Check framing, speed, and intersections with scene geometry.
5. Refine the timing or poses, then preview again.

Use the available camera-motion tools as starting points. Character blocking presets move objects; they do not replace a fully animated character rig.

Produce the output

Check the active camera, frame range, resolution, and output settings before rendering a still or sequence. Inspect the saved result as well as the viewport preview. Sending frames to an AI video workflow produces a separate media result, while camera keyframes remain part of the editable scene.

Export a scene for Three.js

Ask the Agent, “Export this scene as a GLB for Three.js.” It preserves the scene’s object hierarchy, cameras, supported lights, and custom properties, making it easier to continue working with the scene on the web.

Check object placement, camera framing, lighting, and materials in the exported result. Area lights, environment lighting, and some Blender-specific effects need adjustment in the destination. Keep your saved Mixar project as the editable source.

Choose the Agent model

Open the model chip beside Auto in the Agent composer. The picker shows Mixie and your saved provider, with Local when a local model is active. If you have not set up a provider, choose Add a key. The composer’s chip shows who will answer the next request.

Choose an available model in Mixie to use Mixar credits. Expand Thinking on a supported model to choose its reasoning level. Models and thinking options can depend on your plan and the current catalog.

Choose a model that supports images when your request includes visual references. Agent model selection and generation model selection are separate: changing the Agent’s language model does not automatically change the model used for Image to 3D or video.

Switch to your provider or local model

Use the model picker to open AI Provider settings, also available from your profile. Choose a provider and model, enter the required key or sign-in, then Save & Activate. Wait for confirmation before sending a request.

When your provider is active, Agent chat uses that provider and does not spend Mixar credits. Generation jobs remain separate and still use Mixar credits. Local models run inference on your computer but remain part of Mixar’s connected workflow.

To return to Mixar-hosted Agent usage, choose a model in Mixie. A saved API key is paused and stays on this computer; Use my key resumes it without pasting it again. See Use your own AI provider for setup and removal.

What it does

By default the Mixar Agent runs on models Mixar hosts, and Agent requests spend Mixar credits. You can instead point the Agent at your own account with a provider API key, your ChatGPT subscription, or a model running on your own computer. While your own provider is active:

• Agent chat does not spend Mixar credits. Your provider bills you directly for what the Agent uses.
• The model picker shows the active choice. Choose a Mixie model to return to Mixar-hosted Agent usage, or resume your saved provider when you need it.
• Generation jobs are separate. Image, 3D, video and texture generation still run on Mixar and still use Mixar credits. Only the Agent's language model changes.

You set this up once in AI Provider Settings, reachable from your profile menu (top right) or from the composer's model menu. Open it, choose a provider, enter what it asks for, and Save & Activate. Mixar checks the key with the provider once, so a save can take a moment. Wait for the confirmation before treating it as active, and fix any validation error before continuing.

Use a provider API key

For a provider such as Anthropic, OpenAI, Google Gemini or OpenRouter, you need an API key from that provider's own dashboard.

1. Open AI Provider Settings from your profile menu.
2. Choose the provider.
3. Pick a model. Choose a model that can use tools so it can work with your scene.
4. Paste your API key and choose Save & Activate.

Mixar validates the key with the provider, keeps the key on this computer, shows a masked preview, and switches the Agent to your account. To change the key, open the dialog again and enter the new one.

For OpenRouter, enter any model slug from openrouter.ai/models, and choose a model that supports tools.

Use your ChatGPT subscription (Codex)

If you have a ChatGPT subscription, you can run the Agent on it through Codex, OpenAI's command-line sign-in. Mixar uses the sign-in that the Codex CLI creates on your computer. It never sees your password, and your Mixar account is not charged credits while it is active.

1. Install the Codex CLI and run codex login in a terminal. This signs you in and writes a sign-in file on your computer.
2. In AI Provider Settings, choose Codex, pick a model, and Save & Activate. Mixar reads that sign-in file in place; it does not copy your tokens.

Mixar renews the sign-in for you as it nears expiry and keeps it in step with the Codex CLI. If the sign-in expires or is revoked, run codex login again and try your request once more.

Run a model on your own computer

Local model runs the Agent’s model on your own machine. Your requests still use Mixar’s connected services to coordinate Agent work. There are two ways in:

• Managed runtime: choose a curated model and Mixar downloads and runs it for you. A large model can take a few minutes to start.
• Your own server: point Mixar at any OpenAI-compatible server you already run (for example Ollama or llama-server) by giving its base URL, such as http://127.0.0.1:11434, and the model name your server exposes. Add a token only if your server requires one.

Pick a model your server actually provides, then send a small test request — saving a local configuration alone does not prove the model can answer. A local model still works inside Mixar's connected workflow; running it does not make the whole app work offline.

How your key is kept

Mixar never stores your key on its servers. When you save a key, it is kept only on the computer where you entered it, in that computer's secure credential store (the macOS Keychain, Windows Credential Manager or the Linux secret service). Mixar's servers keep only the provider name, the model and a masked preview. Your key is encrypted when sent with an Agent request and used only for that request.

Because the key lives on one computer, it is per computer. If you sign in on another machine, or reinstall, you enter your key there once. For the same reason, a teammate on a shared account does not get your key.

To pause a saved API key, choose a model in Mixie from the composer’s model picker. Agent chat uses Mixar credits again, while the key stays on this computer. Choose Use my key in AI Provider settings to resume it.

To delete an active key, choose Remove key and confirm; for a paused key, choose Forget. Codex and local connections show Disconnect instead. Removing and pausing are different: removing a key means you must enter it again to use it.

If you are asked for your key again

If the Agent says it needs your key to continue, this computer does not currently hold a usable key for your account. This is normal the first time you use a new computer, after reinstalling, or if your account was switched to a different provider somewhere else. Choose Enter key on the message, or open AI Provider Settings, and enter the key for the provider shown. Your request is sent again once the key is in place.

For Codex, "sign-in expired" means you need to run codex login again on that computer. If you would rather not re-enter anything, choose Use Mixar credits instead to run the request on Mixar.

Set up Mixar first

Mixar includes an MCP server. It works with Claude Code, Codex, Claude Desktop, Cursor, VS Code, OpenCode, Windsurf, Cline, Gemini CLI, and any app that can run a local “stdio” MCP server.

1. Open Mixar. You can use free scene tools in guest mode; sign in for paid generation.
2. Open your profile menu (top right) and choose Connect MCP.
3. Pick your app in the list on the left. The dialog shows that app’s exact setup, with the full launcher path filled in, and where it goes. For Claude Code or Codex, click Add to Claude Code or Add to Codex: Mixar adds itself and confirms in the dialog. For another app, click Copy setup and paste it where the dialog says; the Open button names the config file it opens.
4. Optional: tick Let AI apps control Mixar’s interface if the assistant should also see and click Mixar’s menus and buttons. See Work in Mixar from your AI app below.
5. Restart the app’s MCP connection, or restart the app.

Copying or adding a setup enables MCP. Check the status line to see whether AI apps can connect; use Turn on or Turn off beside it to control access. Other app offers a standard JSON setup or launcher command, and Guide opens this page. The sections below explain each app’s setup if you prefer to write it yourself.

Your AI app can start Mixar when it needs to. Use Turn off in the same dialog to stop access; the setup stays available to turn on again. The config holds no account token or API key.

The launcher

Mixar installs one small launcher at a fixed location for your user account:

• macOS and Linux: ~/.mixar/connector/mixar-mcp
• Windows: %USERPROFILE%\.mixar\connector\mixar-mcp.cmd

Most apps do not expand ~, so a config needs the full path, such as /Users/you/.mixar/connector/mixar-mcp. The dialog fills it in for you. When you write a config by hand, replace you with your user name.

The examples below use the macOS and Linux path. On Windows, apps run the launcher through cmd.exe. In a JSON config, that looks like this:

"command": "cmd.exe",
"args": ["/d", "/c", "C:\\Users\\you\\.mixar\\connector\\mixar-mcp.cmd"]

Claude Desktop, Cursor, and similar apps

Select Other app → mcpServers JSON, then Copy setup for the standard JSON most apps use:

{
  "mcpServers": {
    "mixar": {
      "command": "/Users/you/.mixar/connector/mixar-mcp",
      "args": []
    }
  }
}

Claude Desktop, Cursor, Windsurf, Cline, Gemini CLI, and other apps that keep an mcpServers JSON file use it as is. Merge the copied mixar entry into that file’s mcpServers, and keep any servers already there.

• Claude Desktop: choose Settings → Developer → Edit Config.
• Cursor: edit ~/.cursor/mcp.json, or .cursor/mcp.json inside a project.

Any other app

If an app can add a local (stdio) MCP server, give it the launcher path as the command, with no arguments. If it has a tool-call timeout, raise it.

Claude Code and Codex

The quickest way is the dialog: pick the app and click Add to Claude Code or Add to Codex. To set it up yourself instead:

Claude Code

Run this once in a terminal:

claude mcp add --scope user mixar -- /Users/you/.mixar/connector/mixar-mcp

On Windows:

claude mcp add --scope user mixar -- cmd.exe /d /c C:\Users\you\.mixar\connector\mixar-mcp.cmd

Codex

Add this to ~/.codex/config.toml:

[mcp_servers.mixar]
command = "/Users/you/.mixar/connector/mixar-mcp"
args = []
tool_timeout_sec = 610

Keep the longer timeout. Some Mixar calls, such as a large scene script or a render, can take several minutes. Without it, Codex stops waiting after about a minute.

Add to Codex writes only this mixar block, leaves the rest of the file as it was, and keeps the previous file as config.toml.mixar-backup.

VS Code and OpenCode

VS Code

For Copilot agent mode, run MCP: Open User Configuration, or use .vscode/mcp.json in a workspace. Add:

{
  "servers": {
    "mixar": {
      "type": "stdio",
      "command": "/Users/you/.mixar/connector/mixar-mcp",
      "args": []
    }
  }
}

OpenCode

Add this to ~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "mixar": {
      "type": "local",
      "command": ["/Users/you/.mixar/connector/mixar-mcp"],
      "enabled": true,
      "timeout": 60000
    }
  }
}

timeout is how long OpenCode waits for Mixar’s tool list. Mixar may need to start first.

Work in Mixar from your AI app

Ask the assistant to inspect the scene, make a change, and check the result with a viewport image. For example: “Look at my open Mixar scene, add a low wooden bench beside the table, and show me a viewport image.”

Each connection works in one scene tab. The assistant can open a new scene tab for separate work.

Let the assistant use Mixar’s interface

By default the assistant builds, edits, renders, and exports through Mixar’s scene tools, without touching Mixar’s menus and buttons. Tick Let AI apps control Mixar’s interface in profile menu → Connect MCP to let it also see the interface and click, type, and drag in it, then reconnect your app. Your own mouse or keyboard always takes control back. Untick it to stop.

Continue earlier work

Ask the assistant to open a recent project by name, for example “Open my Garden Wedding project and continue.” It can see the projects you recently opened or saved in Mixar. If the open file has unsaved changes, it asks you whether to save or discard them first.

If you quit and reopen Mixar on a saved project, the assistant carries on in the same scene. If you open another file, or the scene wasn’t saved, it reconnects to what Mixar now shows.

Credits

• Inspecting, building and editing scenes and controlling the Mixar UI are free.
• Only AI generation costs Mixar credits, at its usual price.
• The assistant’s own model usage is billed by that app, not by Mixar.

If the connection does not work

• MCP is off or Mixar is unavailable: open Mixar, choose profile menu → Connect MCP, and check the status line. Choose Turn on if it is off, then reconnect your AI app. Sign in when a requested feature requires an account.
• Mixar tools are missing after setup: restart the app’s MCP connection. Check that the config uses the full launcher path, not ~.
• A call timed out: the work may still have finished. Ask the assistant to inspect the scene before repeating anything.
• The assistant says interface control is off: tick Let AI apps control Mixar’s interface in the dialog, then reconnect your app. Scene tools work either way.
• Add to Claude Code fails: Mixar could not find the claude command. Click Copy and run the command in a terminal instead.
• You use a hosted web app: apps that need a remote (OAuth) MCP connection are not supported yet. Use a desktop or command-line app.

Create and revise a tool

1. Open Add-on in the floating island.
2. Describe the tool and where its controls should appear. For example: “Create a sidebar tool that renames selected objects with a prefix and a numbered suffix. Preview the names before applying them.”
3. Read any approval question before the Agent applies a patch. Ask for a correction if the proposed behavior differs from your request.
4. Follow the Agent’s How to use instructions and try the tool on a small saved scene.
5. Ask in the same tab to revise, disable, or roll back the add-on when needed.

The first request prepares the add-on workspace. Add-on Projects Folder in Mixar Preferences controls where new add-ons are created. Keep a saved project before testing a tool that changes scene data. Add-on creation uses the conversation; there is no separate project-creation form.

Reference to finished asset

1. Add a clean product reference to the Moodboard.
2. Generate a mesh in 3D, or ask the Agent to build separate editable parts.
3. Inspect proportions and silhouette from several angles.
4. Refine topology and UVs where needed.
5. Build editable material layers or generate a textured result.
6. Set up a camera and lighting, then render a preview.
7. Save the project, export for your target application, and verify the exported asset there.

Sketch to scene

1. Frame the ground in Zen and open Sketch.
2. Sketch the arrangement and choose Draw to build. Finish with Done and inspect the preview.
3. Specify scale, object relationships, and what must stay editable.
4. Send, answer any questions, and follow the task progress.
5. Review the placement and make focused corrections.
6. Use Cinema Mode to frame a shot; keep the scene and rendered media as separate deliverables.

Essential Blender shortcuts

Use Engine for the native editing workflows below. These tables use Blender’s default keymap with left-click selection. Custom keymaps, Industry Compatible, and Mixar tools such as Sketch or Cinema can change what a key does. Place the pointer over the intended editor and leave text fields before using a shortcut.

Notation: “+” means hold the keys together. A sequence such as G → X → 2 → Enter means press each key in order. MMB means the middle mouse button; Numpad keys are different from the number row.

Navigate and inspect the scene

No numpad or middle button? Use the viewport navigation gizmo and View menu, or search Frame Selected with F3. Preferences provides Emulate Numpad and Emulate 3 Button Mouse. Emulate Numpad repurposes the number row, so it changes how you access the mesh selection shortcuts below. On macOS, Option corresponds to Alt; use the binding shown in the menu for commands with a platform-specific Command shortcut. Function keys may require Fn.

Select, hide, and organize

Local View isolates temporarily; hiding changes visibility. Neither deletes the objects. Linked duplicates share their mesh data, so editing one mesh changes the others. Joining meshes does not automatically weld overlapping vertices or perform a Boolean union.

Transform precisely

Try G → X → 2 → Enter to move along X by 2 current scene units, R → Z → 90 → Enter to turn 90 degrees around Z, or S → 0.5 → Enter to halve the scale. Units and transform orientation affect interpretation; check them before entering dimensions.

Apply and Clear are different: applying scale keeps the visible size while resetting the scale values; clearing scale resets those values and changes the visible size. After Shift + D, right-click cancels only the movement—the new duplicate still exists at the original position.

Edit a mesh

Select a mesh and press Tab to enter Edit Mode. These shortcuts act on mesh components; press Tab again to return to Object Mode.

Extrusion pitfall: right-click after E cancels the movement, but can leave the extruded geometry in place. Use Undo if you meant to cancel the extrusion entirely. Loop cuts depend on suitable edge flow; the shortcut does not guarantee a loop through triangles or poles. Inspect normals on open or non-manifold surfaces even after recalculation.

Save, recover, and repeat

Save before rendering or making large edits. Native Undo and Agent checkpoints are different recovery tools: Undo follows the current scene tab; Undo Whole Document steps across scenes; a checkpoint restore restores an earlier document and attempts to rewind the matching Agent conversation.

Agent, Sketch, and scenes

For the full native reference, see the Blender default keymap. Use Preferences → Keymap to inspect or change your installation’s bindings.

When a result is missing or unexpected

• Generated object is hard to find: locate it in the scene, select it, and frame the selection. Check visibility and scale.
• Reference appears ignored: confirm the selected image and explain its role. Use a vision-capable Agent model for image-bearing requests.
• Generation settings differ from a screenshot: the catalog is dynamic. Use the options currently available for the chosen model.
• Material looks flat: switch to Material Preview or Rendered and check lighting, the active material, UVs, and texture files.
• Scene restored but chat did not rewind: read the scene-only recovery notice. Do not assume the Agent has forgotten the discarded turns; explain the restored scene and restate the next task in full.
• A task is still running: check both the conversation and Queue before resubmitting it.

For account and billing information, visit your dashboard. For help with a reproducible issue, use Report a bug.