Scenario Skills Adds Model Selection and Pricing for AI Agents

What Happened

Scenario Labs released scenario-labs/skills, an open-source skill library that gives AI agents model-selection and pre-execution pricing logic for image, video, audio, and 3D generation. The package exposes these capabilities through MCP (Model Context Protocol), meaning agents connected to the server can query available models, estimate cost before committing to a generation call, and enforce brand consistency rules across outputs. The library ships with integration paths for Blender, Maya, Unreal Engine, and Unity.

Why It Matters

Model selection is currently ad hoc — builders hardcode preferences, write bespoke routing logic, or let cost overruns surface after the fact. Scenario's library abstracts two decisions that every generative pipeline makes repeatedly: which model to invoke for a given task, and what that invocation will cost before it runs. Encoding both as MCP-accessible skills converts them from application-specific glue into reusable primitives that any agent framework can call. The practical beneficiary is the operator managing multi-model fleets, where per-task routing decisions and budget enforcement are recurring costs that don't differentiate the product but do consume engineering time. Brand consistency as a first-class skill parameter is the second-order value: it moves style adherence from a prompt-engineering problem to a policy the agent enforces.

Technical Details

The skills operate over MCP, so any MCP-compatible agent runtime — Claude tooling, custom orchestrators, or IDE-embedded agents — can invoke model discovery, price estimation, and generation requests without a bespoke SDK. Pricing is exposed as a pre-flight query: the agent checks cost against a task specification before spending, enabling budget-aware branching inside a single reasoning loop. Brand consistency is maintained through reference-based conditioning rather than post-hoc filtering, applied across the four modalities. DCC integrations (Blender, Maya, Unreal, Unity) suggest the library targets production pipelines where generated assets land directly in scene graphs rather than in a review queue. The limitations are implicit in the architecture: the price oracle is only as accurate as the provider pricing it reflects, and routing quality depends on how well the model-selection skill is calibrated against actual output quality — a dimension the library exposes but does not itself solve.

Operational Impact

Builders can remove hardcoded model IDs and per-provider cost tables from their pipelines; the agent queries the skill instead, which means model swaps and price changes propagate without code deploys. Budget enforcement shifts left — a generation that would exceed a task's allocation can be rejected or rerouted before the API call fires, eliminating the current pattern of discovering overruns in monthly billing. For teams running DCC-heavy workflows, assets arrive in-engine with brand policy already applied, collapsing a review-and-rework loop. The workflow that becomes obsolete is the manual model-picker spreadsheet: operators maintaining per-shot model preferences by hand now have a callable interface for the same decision. What gets harder is model governance — once selection is automated, auditing why a specific model was chosen for a specific asset requires logging the skill's decision trace, a practice most pipelines don't yet have.