MCP server
Cinevva runs a Model Context Protocol server, so an AI agent can rig models, search assets, and apply animations as tools rather than by writing HTTP calls.
https://api.cinevva.com/create/mcpTransport is Streamable HTTP. Authentication is OAuth 2.0 with dynamic client registration, so a compliant client discovers and registers itself with no manual key exchange — you sign in once in a browser window and the client holds the token from then on.
Connecting
In Claude Code:
claude mcp add --transport http cinevva https://api.cinevva.com/create/mcpIn Claude Desktop or any other MCP client, add the server URL in the client's connector settings. The client will open a browser for the OAuth consent step. The server is also listed in the public MCP registry as com.cinevva/game-creator.
What the agent gets
The tools most relevant outside Cinevva itself run server-side with no browser tab, so an agent can use them headlessly:
| Tool | What it does |
|---|---|
rig_model | Auto-rig a mesh. Returns a durable task handle; spends credits |
apply_animation | Resolve clips and return the code to play them |
list_animations | Search the 260-clip CC0 catalog |
search_assets | Search the CC0 asset library and live providers |
list_asset_packs | Enumerate available packs |
import_asset | Pull an asset in, including Freesound audio and Poly Haven HDRIs |
search_site_content / read_site_page | Search Cinevva's own docs and guides |
search_library_docs / read_library_docs | Look up bundled library documentation |
rig_model is asynchronous. With a client that supports the MCP tasks extension it returns a task handle the agent can await; otherwise it returns a job id to poll.
You do not need a Cinevva game to use these. Pass gameId: "external" and the asset and rigging tools work against whatever project you are building.
The remaining tools in the server are for driving a live Cinevva game session — reading its console, executing JavaScript in it, taking screenshots, simulating input — and those do require an open browser tab, because there is a running game for them to talk to.
Billing
MCP calls bill exactly like the HTTP API: a rig through rig_model costs the same as a rig through POST /v1/rigs, and browsing catalogs is free. The same subscription requirement applies.
Rigs started by an agent appear in your account alongside everything else, so GET /v1/account is still the place to watch usage.
When to use MCP instead of the API
Use MCP when an AI agent is the thing doing the work and you want it to discover the capability rather than be told about it. The tool descriptions carry the constraints — that rig_model spends credits, that apply_animation does not — so a well-behaved agent makes sensible choices without you writing the guardrails.
Use the HTTP API when your own code is in control: batch pipelines, build steps, or a backend serving your own users. Deterministic call sites belong in code, not in a model's judgement.