Skip to content
solomotion
AgentsWhat agents can do

Every command is a tool

The studio runs on one command registry. The buttons you click call it, scripts call it, and an agent calls it over MCP. So an agent can do what you can do, and nothing is a separate, thinner API.

The Agents panel in the left rail, with the line An agent that designs, animates and edits with you, and two buttons: Connect a model and Use your agent app.

How an agent works

  1. Orient. doc_overview lists comps, sizes and the selection. layers_list reads one comp.
  2. Build. One api_batch per idea, or one script_run for anything repetitive.
  3. Look. view_render returns a frame as an image; view_frames returns several across time.
  4. Refine what it saw, then report in a few sentences.

The server also serves the agent skill as resources (solomotion://skill/SKILL.md and its references): how to work well in the studio, plus a default design system for type, spacing and timing.

The command map

JobCommands
Readdoc_overview, layers_list, layer_get, layer_properties, media_list
Learn a shapeapi_describe, catalog_list, catalog_get
Make layerslayer_create (text, shape, SVG path, media, nested comp, adjustment, waveform, trace)
Change valueslayer_set, layer_pivot, text_set, text_style, text_outline, layer_patch
Motionanimation_set, keyframes_set, keyframes_clear, keyframes_inspect, keyframes_simplify, expression_set
Effects and coloureffects_add, effects_update, effects_look, effects_morph, grade_node, grade_effect, grade_read
Structurelayers_group, layout_set, layers_nest, layers_restack, layers_duplicate, layers_copy, layers_delete
Timetimeline_move, timeline_trim, timeline_split, timeline_cut_ranges, timeline_close_gaps, transition_add, audio_detach, audio_rejoin
Comps and pagesframe_create, frame_duplicate, frame_update, frame_delete, page_create, page_update, page_delete
3Dworld_create, world_add, world_extrude, world_material, world_look, world_lattice, world_sculpt, world_line_art, world_import_model, world_move_into
Characterscharacter_detect, character_bringToLife, rig_*, poses_*, behaviors_*, shapekey_*, speech_*, machine_*
Physicsphysics_body, physics_force, physics_emit, physics_joint, physics_scene, physics_inspect
Trackingtrack_create, track_analyze, track_correct, track_follow, track_scene
Flipbooksflipbook_cells, flipbook_edit
Scriptsscript_run, script_save, script_read, script_reify
Seeview_render, view_frames
Export and saveexport_render, export_wait, project_save, project_import, project_open
Undohistory_undo (its own steps only)

The list your agent sees is the live one. A field or effect that ships next month shows up in api_describe, layer_properties and catalog_list the day it lands.

Three levels of writing

Agents use the highest level that fits:

  1. Verbs such as layer_create or animation_set. They keep the rules a raw edit would break.
  2. layer_set by property path, for any animatable value: x, opacity, fill.<id>.color. Once a property has keys, a write adds a key.
  3. layer_patch for everything else, validated against the schema. An invalid patch changes nothing and names the wrong field.

When it cannot

The agent says so, does what it can, and files report_issue with what it tried. A command that needs a model provider you have not connected fails with a hint naming what is missing.