Best MCP for game assets in Claude Code

Why one connection rather than several

A game needs assets from services that have nothing to do with each other: an image model for concept art, a different one for 3D, another for speech, another for music. Wiring each of those into Claude Code separately means a server per provider, a key per provider, and a bill per provider. A catalog collapses that into one connection, one balance, and one place where prices are visible before anything runs.

The practical difference shows up in a single request. "Make me a character sheet, then a 3D model of it, then a line of dialogue" touches three services; with one catalog the agent lists all three with prices and runs them in order, and the files land in the same folder. With three separate servers it is three conversations, three approvals, and three sets of credentials to keep out of your repository.

The three routes

One thing people expect to differ here does not. What puts finished files in your project is the Actionway Skill, and both the plugin and the CLI install it — the plugin bundles it alongside the Remote MCP connection. Its rule is identical in the two: when the host can write files, download the assets into an outputs/ folder; when it cannot, hand back the URLs. So in Claude Code or Codex a model, a sheet, and a voice line all arrive where your build script can find them, on either route. The exception is a bare Remote MCP connector added on its own, with no Skill behind it — nothing then tells the agent to save anything, so you may get URLs on a machine that could have written the files.

Setting up the CLI route

npm install --global @actionway/cli@latest
actionway init
mkdir -p "$HOME/.claude/skills"
cp -R "$HOME/.actionway/skill/actionway" "$HOME/.claude/skills/actionway"

That last copy step is the one people miss. init authenticates you and stages the Skill, but it deliberately does not write into an agent's own configuration directory — so until you copy the folder, Claude Code does not know the catalog exists. You need Node.js 20 or newer and an Actionway account.

Setting up the other two

claude plugin marketplace add PawLogic/actionway-plugins
claude plugin install actionway@pawlogic

The Remote MCP route needs no commands at all: add a custom connector pointing at https://mcp.actionway.ai/mcp, approve the OAuth prompt in your browser, and the catalog appears in the client's tool list. That is also the route for clients that cannot run local commands — the chat apps, the mobile apps, and anything corporate-managed. The get started page lists the exact menu path per client, because every client names these settings differently.

What you get on any route

1. A catalog the agent can search

   The agent looks up what fits your request rather than being told which tool to call. Searching is free; nothing is charged until you pick a priced option.

2. A price before every run

   The first reply lists the services that fit with their prices — an image at $0.15 per run, a 3D model from $0.22 per run (none-standard-triangle), a music run at $0.20 per run. Picking one is the approval.

3. One balance for everything

   No subscription and no per-provider account. Every run is itemised, and an empty wallet stops a job before anything is charged.

4. The same catalog in Codex

   Codex users copy the same Skill folder into ~/.codex/skills/actionway, or use the matching plugin commands. Same services, same prices, same wallet.

Which one to pick

Pick Remote MCP if you cannot install software, if you move between machines, or if a connector is the only thing your client accepts. Pick the plugin if you already manage plugins for your team and want the connection to arrive the same way everything else does — it is the shortest install that still brings the Skill with it. Pick the CLI if you want the actionway command itself: something you can run in a plain shell, put in a script, or use to check a balance without opening an agent at all.

You can also change your mind later without losing anything: the balance, the run history, and the catalog are account-level, so switching routes is a setup change rather than a migration. One difference does follow you, though. Every guide in this series ends with a local step — trimming a music bed, mixing audio under a clip, converting a table to CSV, checking a mesh's polygon count — and that step needs an agent that can run commands on your machine. In a terminal agent any route gives you that; in a chat client, Remote MCP gives you the generation but the finishing work moves elsewhere.

Trying it on something small

If that reply lists services and prices, the connection works. From there the game asset guide walks one character through concept art, store art, a 3D model, and a voice line, and the 2D-to-3D comparison covers the one setting that decides whether a generated model is usable.