atlas-cloud-media – DevOps & Infrastructure skill for AI agents

Atlas Cloud Media

Overview

Use Atlas Cloud's asynchronous media API to generate images or videos. This

source-only skill describes model discovery, schema validation, task

submission, bounded polling, and safe output retrieval; it does not bundle an

SDK, executable, or hosted runtime.

When to Use This Skill

• Use when the user explicitly asks to generate an image or video with Atlas

Cloud.

• Use when an existing workflow needs an Atlas Cloud image or video generation

request and can make HTTPS calls.

• Use when model-specific parameters must be discovered before submission.
• Do not use this skill for OpenAI-compatible text chat; that API has a

different base URL and contract.

Preconditions

1. Confirm the user is authorized to send the prompt and any reference media

to a third-party service.

2. Explain that generation is paid and obtain approval before submitting a

billable request.

3. Require ATLASCLOUD_API_KEY to be present in the environment. Never ask the

user to paste it into chat, source files, command history, or logs.

4. Confirm the output directory and whether the user wants image generation,

video generation, or both.

API Contract

| Operation | Method and endpoint |

| --- | --- |

| List models | GET https://api.atlascloud.ai/api/v1/models |

| Generate image | POST https://api.atlascloud.ai/api/v1/model/generateImage |

| Generate video | POST https://api.atlascloud.ai/api/v1/model/generateVideo |

| Poll task | GET https://api.atlascloud.ai/api/v1/model/prediction/{id} |

Generation and polling requests use these headers:

Authorization: Bearer $ATLASCLOUD_API_KEY
Content-Type: application/json

The model catalog is public. Each catalog entry includes a schema URL; fetch

that schema and validate parameters against it before sending a paid request.

Do not guess parameters from another model, because names such as size,

ratio, aspect_ratio, image, and image_url are model-specific.

Workflow

0. Create a Private Per-Run Workspace

Run the remaining shell snippets in the same shell session. Create a private

directory before writing prompts, responses, prediction IDs, or signed URLs;

the parameter expansion in later steps fails closed when this setup was skipped.

umask 077
atlas_tmp_dir=$(mktemp -d "${TMPDIR:-/tmp}/atlas-cloud-media.XXXXXXXX") || exit 1
chmod 700 -- "$atlas_tmp_dir"
trap 'rm -rf -- "$atlas_tmp_dir"' EXIT

1. Discover and Validate a Model

Fetch the catalog, filter by type (Image or Video), and match the user's

requested capability. Read the selected entry's schema, verify that all

required fields are present, and show the model and billable action to the user

before submission.

Example discovery request:

curl --fail --silent --show-error --connect-timeout 10 --max-time 60 \
  "https://api.atlascloud.ai/api/v1/models" \
  --output "${atlas_tmp_dir:?run private workspace setup first}/models.json"

jq -r '.data[] | select(.type == "Image") | [.model, .displayName, .schema] | @tsv' \
  "$atlas_tmp_dir/models.json"

2. Submit One Generation Task

Build the JSON body in a file so that quoting is deterministic and request

details can be reviewed without exposing the API key.

Image example using a catalog-confirmed model:

jq -n \
  --arg model "qwen-image-3.0/text-to-image" \
  --arg prompt "A paper-cut city map in blue and white, clean editorial style" \
  '{model: $model, prompt: $prompt, size: "1024*1024", n: 1}' \
  > "${atlas_tmp_dir:?run private workspace setup first}/request.json"

curl --fail --silent --show-error --connect-timeout 10 --max-time 60 \
  --request POST \
  "https://api.atlascloud.ai/api/v1/model/generateImage" \
  --header @- \
  --header "Content-Type: application/json" \
  --data @"$atlas_tmp_dir/request.json" \
  --output "$atlas_tmp_dir/submit.json" <<EOF
Authorization: Bearer ${ATLASCLOUD_API_KEY}
EOF

Video example using a catalog-confirmed model:

jq -n \
  --arg model "bytedance/seedance-2.0-fast/text-to-video" \
  --arg prompt "A small paper boat crossing a calm pond, locked camera" \
  '{
    model: $model,
    prompt: $prompt,
    duration: 4,
    resolution: "480p",
    ratio: "16:9",
    generate_audio: false,
    watermark: false
  }' > "${atlas_tmp_dir:?run private workspace setup first}/request.json"

curl --fail --silent --show-error --connect-timeout 10 --max-time 60 \
  --request POST \
  "https://api.atlascloud.ai/api/v1/model/generateVideo" \
  --header @- \
  --header "Content-Type: application/json" \
  --data @"$atlas_tmp_dir/request.json" \
  --output "$atlas_tmp_dir/submit.json" <<EOF
Authorization: Bearer ${ATLASCLOUD_API_KEY}
EOF

Check that .data.id is a non-empty string before polling. Treat a non-2xx

response or a missing ID as submission failure; do not retry a billable request

automatically because the original task may still have been accepted.

3. Poll with a Deadline

Poll every three seconds. Accept completed or succeeded as success, stop on

failed or timeout, and enforce a ten-minute wall-clock deadline in the host. The attempt cap below is not itself a ten-minute bound because HTTP request time also counts. Preserve the prediction ID

for diagnostics, but never log request headers or the API key.

prediction_id=$(jq -er '.data.id | select(type == "string" and length > 0)' \
  "${atlas_tmp_dir:?run private workspace setup first}/submit.json")

for attempt in $(seq 1 200); do
  sleep 3
  curl --fail --silent --show-error --connect-timeout 10 --max-time 60 \
    "https://api.atlascloud.ai/api/v1/model/prediction/$prediction_id" \
    --header @- \
    --output "$atlas_tmp_dir/prediction.json" <<EOF
Authorization: Bearer ${ATLASCLOUD_API_KEY}
EOF

  status=$(jq -r '.data.status // "unknown"' "$atlas_tmp_dir/prediction.json")
  case "$status" in
    completed|succeeded) break ;;
    failed|timeout)
      jq -r '.data.error // "Atlas Cloud generation failed"' \
        "$atlas_tmp_dir/prediction.json" >&2
      exit 1
      ;;
  esac
done

test "$status" = "completed" || test "$status" = "succeeded"

4. Download and Verify the Output

Read the first HTTPS URL from .data.outputs. Atlas output URLs are temporary,

so download promptly. Validate each redirect host and resolved IP against approved public media destinations; HTTPS alone does not prevent access to local/private networks. Do not send Authorization or any other Atlas request

headers to the output host. Reject non-HTTPS URLs and inspect the downloaded

file's content type and size before treating it as a valid deliverable.

output_url=$(jq -er '.data.outputs[0] | select(startswith("https://"))' \
  "${atlas_tmp_dir:?run private workspace setup first}/prediction.json")

curl --fail --silent --show-error --connect-timeout 10 --max-time 60 --location --max-redirs 3 --proto =https --proto-redir =https \
  "$output_url" \
  --output "$atlas_tmp_dir/output.bin"

test -s "$atlas_tmp_dir/output.bin"
file "$atlas_tmp_dir/output.bin"

# ATLAS_OUTPUT_DIR must be the user-approved destination. Resolve it to a
# physical directory, copy into an exclusive same-directory temporary file,
# then create the final name with one atomic hard-link operation. `ln` fails if
# any target already exists, including a dangling symlink.
atlas_output_dir=$(cd -- "${ATLAS_OUTPUT_DIR:?set the approved output directory}" && pwd -P) || exit 1
atlas_output_path="$atlas_output_dir/atlas-output.bin"
if ! (
  set -eu
  umask 077
  atlas_publish_tmp=$(mktemp "$atlas_output_dir/.atlas-output.XXXXXXXX")
  trap 'rm -f -- "$atlas_publish_tmp"' EXIT
  cp -- "$atlas_tmp_dir/output.bin" "$atlas_publish_tmp"
  chmod 644 -- "$atlas_publish_tmp"
  ln -- "$atlas_publish_tmp" "$atlas_output_path"
); then
  printf '%s\n' "Refusing to overwrite or redirect $atlas_output_path" >&2
  exit 1
fi

Rename the file only after its detected type is known. Report the local path,

model ID, dimensions or duration, and whether the output passed basic playback

or decode validation.

Failure Handling

• 401 or 403: stop and ask the user to verify access. Do not print or rotate

the key automatically.

• 400 or 422: fetch the model's current schema and correct the payload. Do

not blindly resubmit.

• 429: stop and report rate limiting; respect any Retry-After value.
• 5xx or network timeout: first poll a known prediction ID. Do not create a

second paid task unless the user approves the possible duplicate charge.

• failed or timeout: report the sanitized service error and prediction ID;

do not claim an output was generated.

• Missing or invalid media: keep the original response for diagnosis, do not

overwrite an existing destination, and do not mark the task complete.

Best Practices

• Use the public catalog and per-model schema immediately before generation.
• Keep request and response artifacts in one private per-run directory and let

the exit trap

… (full content is in the download package)