Generate Images with MuAPI

Generate Images with MuAPI

Use this skill when an agent needs hosted image generation through MuAPI's unified API. First inspect the live catalog and select an exact image model suited to the task; the catalog includes multiple FLUX, Nano Banana, Seedream, GPT Image, Midjourney, Qwen, and other model families. The workflow then makes one explicitly authorized generation request, polls the returned prediction, and downloads the completed image without forwarding the API key to the output host. Do not assume that a model name or payload is permanent: read the current model schema before adding model-specific fields. This skill is for image workflows, not video, audio, 3D, or chat requests.

Installation

Manual installation

git clone https://github.com/agentskillexchange/skills.git
cp -R skills/skills/generate-images-with-muapi ~/.agent-skills/generate-images-with-muapi

Optional third-party installer

The skills npm package is maintained by Vercel Labs / third parties. If you use it, pin the package version:

npm exec --package=skills@1.5.7 -- skills add agentskillexchange/skills --skill generate-images-with-muapi

Requirements and authorization

• Bash, curl, jq, and file
• A MuAPI key exported as MUAPI_API_KEY
• An explicit user-approved prompt, output path, and potentially billable request
• A finite polling budget

Never print or commit the key, put it in a command argument, or enable shell tracing. Treat a generation POST as accepted if its outcome is ambiguous; do not retry it automatically.

1. Discover the current model and request contract

Read the live catalog immediately before building a request. The catalog response currently uses top-level models and total fields. Each model entry includes fields such as name, category, and an already-versioned endpoint path. Choose an exact text-to-image or image-to-image model; do not assume the catalog has a .data[] array, a .type field, or an inline .schema URL:

curl -fsS --max-time 30 \
  https://api.muapi.ai/api/v1/models \
  -o /tmp/muapi-models.json

jq -r '
  .models[]
  | select((.category // "" | ascii_downcase) | test("^(text to image|image to image)$"))
  | [.name, (.category // ""), (.endpoint // "")] | @tsv
' /tmp/muapi-models.json

Use the endpoint returned by the catalog, not a remembered alias. Payload fields differ by model; the minimal prompt payload below is only a starting point. Add size, reference-image, or quality fields only when the selected model's current model documentation or request contract supports them.

2. Build and review the request

Construct JSON with jq so prompt text is escaped. Replace MODEL_ENDPOINT and add only fields validated against that model's current request contract:

API_ORIGIN='https://api.muapi.ai'
MODEL_ENDPOINT='/api/v1/replace-with-the-exact-endpoint-from-the-catalog'
PROMPT='a small red fox reading a book beneath a lantern, storybook illustration'
case "$MODEL_ENDPOINT" in
  /api/v1/*) ;;
  *) echo "Refusing an endpoint that is not an /api/v1 catalog path" >&2; exit 1 ;;
esac
SUBMIT_URL="${API_ORIGIN}${MODEL_ENDPOINT}"
jq -n --arg prompt "$PROMPT" \
  '{prompt: $prompt}' \
  > /tmp/muapi-image-request.json
jq . /tmp/muapi-image-request.json

3. Submit exactly once

Use the documented model endpoint and do not add curl --retry or a POST retry loop:

curl -fsS --max-time 60 \
  -X POST "$SUBMIT_URL" \
  -H "x-api-key: $MUAPI_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json' \
  --data-binary @/tmp/muapi-image-request.json \
  -o /tmp/muapi-image-submission.json

REQUEST_ID=$(jq -er '.request_id // .data.request_id // .id // .data.id' \
  /tmp/muapi-image-submission.json)
printf 'request_id=%s\n' "$REQUEST_ID"

Stop if the request times out or returns no request ID. Reconcile an ambiguous request through the MuAPI account or support rather than creating a second generation job.

4. Poll with a finite budget

Prediction GET requests may be retried because they do not create replacement jobs. Poll at most 60 times, stop on terminal failure, and preserve the request ID if the budget is exhausted:

PREDICTION_URL="https://api.muapi.ai/api/v1/predictions/$REQUEST_ID/result"
for attempt in $(seq 1 60); do
  curl -fsS --max-time 30 \
    -H "x-api-key: $MUAPI_API_KEY" \
    -H 'Accept: application/json' \
    "$PREDICTION_URL" -o /tmp/muapi-image-result.json || true

  STATUS=$(jq -r '.status // .data.status // empty' /tmp/muapi-image-result.json 2>/dev/null || true)
  case "$STATUS" in
    completed) break ;;
    failed|timeout|canceled|cancelled)
      jq -r '.error // .data.error // ("prediction ended with status " + (.status // .data.status // "unknown"))' \
        /tmp/muapi-image-result.json >&2
      exit 1
      ;;
  esac

  if [[ "$attempt" -eq 60 ]]; then
    echo "Polling budget exhausted; preserve request ID $REQUEST_ID" >&2
    exit 1
  fi
  sleep 2
done

5. Download and inspect the artifact

Read the first output URL only after completion. Require HTTPS and do not send MUAPI_API_KEY to the returned CDN URL:

OUTPUT_URL=$(jq -er '.outputs[0] // .data.outputs[0]' /tmp/muapi-image-result.json)
case "$OUTPUT_URL" in
  https://*) ;;
  *) echo "Refusing non-HTTPS output URL" >&2; exit 1 ;;
esac

OUTPUT_PATH='./muapi-output.png'
curl -fsS --max-time 60 --proto '=https' --max-filesize 26214400 \
  "$OUTPUT_URL" -o "$OUTPUT_PATH"
file --mime-type "$OUTPUT_PATH"

Open or otherwise inspect the file before reporting success. Check that it is a valid image and meets the prompt, composition, safety, privacy, likeness, trademark, and usage-rights constraints.

Official references

• MuAPI model catalog
• MuAPI API reference
• MuAPI image models
• MuAPI access keys