Intent Tools¶

Intent tools translate a plain-language request into deterministic Unity tool calls. They are useful when the desired result is clear but the exact object or tool sequence is not. For known paths and exact changes, prefer the typed tool or batch: it is easier to review and reproduce.

Intent tools are direct calls; they cannot run inside batch.

Tool	Use it for	Mutates the project?
`ask`	Answer a read-only question about the current scene	No
`do`	Plan and apply a scene change from a broad request	Yes
`animator_intent`	Build Animator parameters, states, and transitions	Yes
`vfx_intent`	Configure a Particle System from a preset or description	Yes
`ui_intent`	Build a Canvas-based UI hierarchy	Yes
`uitk_intent`	Build a UI Toolkit UXML/USS panel	Yes

Generated plans are validated, but they still operate on your project. Preview ambiguous or wide changes and verify the result with a read tool.

`ask(question)`¶

Use ask for scene questions that would otherwise require several reads.

answer = await ask("Which enemies have a Health component?")

The router gathers read-only scene context and may use configured LLM sampling to summarize a complex result. Questions that look mutating are rejected; use do or a typed mutation tool instead.

`do(intent, dry_run=False)`¶

Use do when the request is broad or the necessary scene paths are not known. Preview first when the operation could affect several objects:

plan = await do(
    "Create three evenly spaced spawn points under /Level/Spawns",
    dry_run=True,
)

# Review the returned DSL. To execute those exact commands, copy the plan after
# the "DRY RUN plan:" line into a direct batch call.
result = await batch(commands="""
create_object name=Spawn1 parent=/Level/Spawns
set_property path=/Level/Spawns/Spawn1 component=Transform prop=position value=-2,0,0
create_object name=Spawn2 parent=/Level/Spawns
set_property path=/Level/Spawns/Spawn2 component=Transform prop=position value=0,0,0
create_object name=Spawn3 parent=/Level/Spawns
set_property path=/Level/Spawns/Spawn3 component=Transform prop=position value=2,0,0
""", on_error="stop")

do reads a compact hierarchy, generates a batch plan, validates the plan, and then executes it. A second do call samples a new plan; a dry run is advisory, not an approval token for the next call. Execution uses on_error="continue". After a small partial failure, do may sample and run one repair plan, while successful earlier operations remain applied. An invalid or unavailable initial plan fails without applying commands.

When every required command is in its Unity-Undo allowlist, use the guarded scene-change workflow for explicit stop-on-error, verification, and save gates.

`animator_intent(target, intent, dry_run=False)`¶

Creates Animator parameters, states, a default state, and transitions from a description.

plan = await animator_intent(
    target="/Player",
    intent="Idle and Walk states controlled by a Speed float at threshold 0.1",
    dry_run=True,
)

The generated DSL supports float, int, bool, and trigger parameters; states with animation clips; a default state; and transition conditions. Confirm that named clips exist before applying the plan. Use the typed animator tool when state names and transitions are already known.

`vfx_intent(target, intent, kind="auto", dry_run=False)`¶

Configures a Particle System. kind currently accepts auto or particle; shader/material intent is not implemented by this tool.

Five exact preset names bypass LLM generation:

fire_explosion
magic_burst
dissolve
glow_outline
smoke_trail

# Deterministic preset.
result = await vfx_intent("/Effects/Explosion", "fire_explosion")

# Preview a generated particle plan.
plan = await vfx_intent(
    target="/Effects/Dust",
    intent="A subtle dust puff that fades quickly",
    kind="particle",
    dry_run=True,
)

The target must already contain the Particle System that the generated particle commands configure.

`ui_intent(intent, parent=None, template=None, dry_run=False)`¶

Builds Canvas-based (uGUI) elements. It does not create UI Toolkit UXML or USS; use uitk_intent for that.

The deterministic templates are hud, menu, dialog, and grid:

plan = await ui_intent(
    "Create the standard heads-up display",
    parent="/UI",
    template="hud",
    dry_run=True,
)

Without a template, configured LLM sampling generates a uGUI hierarchy plan:

plan = await ui_intent(
    "Create a centered pause panel with Resume and Quit buttons",
    parent="/UI/Canvas",
    dry_run=True,
)

See UI authoring for typed uGUI and UI Toolkit workflows.

Sampling and budget status¶

Generated intent plans require optional LLM sampling to be enabled and the supported Claude CLI backend to be installed and authenticated. Deterministic VFX presets and UI templates do not make a generation call. Configure sampling under MCP > Settings > LLM Sampling; see Settings.

budget_status() reports the current session/day estimate and any features skipped by the budget router. Treat it as an estimate, not a provider invoice. Limits and model pricing can change, so this guide intentionally does not hard-code a per-request price.

Reliable workflow¶

Use a typed tool when the target and operation are known.
Otherwise, call the intent tool with dry_run=True when it supports preview.
Check target paths, asset names, and the breadth of the returned plan.
Apply the request.
Verify with inspect, get_hierarchy, an Animator read, or a screenshot as appropriate.

Intent Tools¶

ask(question)¶

do(intent, dry_run=False)¶

animator_intent(target, intent, dry_run=False)¶

vfx_intent(target, intent, kind="auto", dry_run=False)¶

ui_intent(intent, parent=None, template=None, dry_run=False)¶