An OpenRouter for what GPT-6 Astra can do

{
  "ok": true,
  "status": "matched",
  "candidates": [
    {
      "toolRef": "media.image.generate",
      "title": "Generate Image",
      "suggestedVariants": [
        "banana-2",
        "banana-pro",
        "gpt-image-2",
        "seedream",
        "seedream-5-lte"
      ],
      "pricing": {
        "status": "priced"
      }
    }
  ]
}

That is a real search result, trimmed only for length. Nothing ran, nothing was charged, and the agent already knows which capability fits, which models it offers, and what each costs — everything it needs to make the first reply a priced plan. Discovery is the part of "skills" people underrate: an agent with one tool can do one thing, and an agent that can look things up can do whatever the capability catalog holds.

What happened in between

OpenAI released GPT-6 Astra on September 3, 2026 around computer use and long-running agent tasks, and Codex CLI 0.153.4 made it the bundled default model. A Codex Skill is a folder of instructions and references the agent loads when a request matches its description. The Actionway Skill's description covers media generation, media understanding, downloads, specialized search, and live data, so a request in any of those areas pulls it in without you naming it.

Inside the Skill is a capability index — every enumerated capability, its models, and indicative prices — plus one reference file per capability. Most requests resolve from that index with no call at all. The long tail (social platform statistics, market data families, public datasets) is not enumerated; the agent reaches it with the free search you saw above, and only says something is unavailable after it has actually searched. Either way the first reply lists two or three priced options, neutrally — no default, no recommendation; say "you pick" and it takes the best-quality tier. Picking one is the approval, and only runs the platform flags for extra confirmation pause for another question.

Pick a model, pay per run

The poster brief from the case would land on the image generation capability: GPT Image 2 at $0.15 per run, Seedream from $0.10 per run, the Banana models from $0.20 per run. The same Skill routes a video brief to the video generation models, a voiceover to voice generation, a ticker to stock and company data, and a TikTok link to search and source. Prices are rendered from the same registry the models library shows, so what you read here is what the agent quotes.

Asynchronous work — video, music, 3D — runs as a job the agent waits on; synchronous calls such as quotes and searches return in the same turn. Finished files land in an outputs/ folder in your project with descriptive names. If the wallet cannot cover a run, the agent stops before anything is charged and gives you a top-up link with the amount prefilled. There is no subscription; one balance covers every capability, and the wallet statement itemizes each call.

What the Skill is, and what it is not

It is a folder of instructions and reference files Codex reads when a request matches its description, plus the CLI that runs the calls. It is not a plugin marketplace, not a key manager, and not a wrapper around one provider. There are no API keys to create per model — every run is billed to one Actionway wallet — and there is no tool to write per capability, because the catalog is the tool. When a new model lands in the catalog, the index inside the Skill updates with the CLI, and the agent offers it the next time a request fits.

Two habits make it work well. First, describe the outcome, not the service: "a poster for the launch" lets the agent choose between image models and show you the prices, while "call GPT Image 2" skips that and runs one model. Both are fine; the first one tells you what else you could have picked. Second, keep the outputs folder: every file the agent brings back lands there with a descriptive name, and the next request can point at any of them as a reference, which is how a poster becomes a clip and a clip gets a voiceover in the same session.

Set it up once

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

You need Node.js 20 or newer and an Actionway account; after sign-in the agent reports your wallet balance and, if it is empty, links you straight to top-up. Prefer no local CLI? codex plugin marketplace add PawLogic/actionway-plugins then codex plugin add actionway@pawlogic loads the same Skill over Remote MCP. The get started page lists every route per client side by side, and the video guide shows a full paid run end to end.