Sushibelt is a harness for game-asset AI: it turns a pile of models into a pipeline you can run, pay for, remember, and hand to another agent. This page is the part you reach without the UI — what exists today, and what is on the way.
Key-authenticated, over HTTP. Trigger a run from CI, a build script, a Unity plugin or an agent. Same credits, same audit trail as a run started on the canvas — nothing about billing changes because you skipped the UI.
A build script can regenerate placeholder art on every commit. A nightly job can produce a hundred variants. An agent can call it as a tool and go and get its own assets, without a human picking models in between.
curl -X POST \
https://app.sushibelt.ai/api/v1/workflows/$WORKFLOW_ID/runs \
-H "Authorization: Bearer sk_live_..." \
-d '{ "inputs": { "prompt": "weathered iron lantern" } }'Read top-down for what you get, bottom-up for what runs. Every layer says which parts exist today and which are planned — we do not blur that line.
The product used without its UI. Claude Code, your own pipeline, a Unity plugin — calling Sushibelt’s tools with a team API key. Same credits, same audit.
A public, key-authenticated run API exists for workflows.
An MCP server exposing the tool registry verbatim, plus resources for the library, the catalogue and model pages. Deliberately last in the sequence: it is the same layer over a different transport.
“Make me a rigged fox in three styles and show me the best.” Planning and coordinating several runs toward a goal.
The broker and scheduler run one graph at a time.
A loop that plans a set of runs, executes them, scores results against the goal, retries a failed stage with a different model, and stops at the budget you set. Every step is a real workflow run with its own charges — a graph of graphs — so nothing about billing changes.
Tools are the platform’s verbs an AI can call. Memory is what it knows about your team without being told again.
The in-app assistant has three tools — interpret a brief, propose a workflow, offer options — and reads the measured model inventory. One conversation per agent; every run and asset is recorded.
One typed tool registry with an audit table, callable by workflow nodes. A memories table per team — “we ship low-poly”, “always rig characters” — written only when the assistant says it is remembering, visible and deletable on one page.
“The orc from last month.” “Something like this concept.” “A template that makes rigged low-poly creatures.”
The model inventory is tagged and ranked; Discover has same-model and same-type similarity.
Vector search across your library, the public catalogue, starter templates and the model inventory — by text and by image. Embedding runs as a charged step like any other node.
The runtime around a model — sequencing, tools, retries, limits, observability. What makes a capable-but-raw model usable at work.
A workflow is a graph compiled and run by the broker. Each step is routed to a provider, charged in credits, retried on failure and recorded. Conditionals and iteration gates exist; runs can be scheduled off-peak; every run is callable over HTTP.
An agentic loop on top, where a model decides the next step at run time — see Orchestration.
Image, texture, mesh, rig, audio — from the majors and from vendors on the marketplace.
Every provider on the line is charged from one credit balance. The model inventory is tagged and ranked on real runs, not on vendor claims.