DreamLayer image workflows

Use this when a task needs a finished image file rather than a description of one, and the result has to land in the user's working directory: a product shot on a clean background, a logo direction to review, a set of campaign variations, a transparent cutout, or an animated sprite sheet for a game project.

Setup

Node.js 22.12 or later, and a DreamLayer API key from platform.dreamlayer.io.

{"mcpServers":{"dreamlayer":{"command":"npx","args":["-y","@dreamlayer/mcp@0.4.0-beta.5"],"env":{"DREAMLAYER_API_KEY":"YOUR_KEY"},"connectionTimeoutMs":90000}}}

Keep the timeout. On a machine that has never run this package, npx -y downloads it before the server can answer. Two clean profiles measured that first start at 13 and 50 seconds; OpenClaw waits 5 seconds by default and reports the server as failed, so the first openclaw mcp add or probe fails without connectionTimeoutMs. Hosts that use a different key for the same idea need the equivalent raised. Later launches are fast, and npm i -g @dreamlayer/mcp@0.4.0-beta.5 beforehand avoids the wait entirely.

Supply the key through the host's secret mechanism. Never place it in a prompt, a commit or a skill file. Without it the server exits and says which variable is missing.

The server exposes seven tools: dreamlayer_capabilities, dreamlayer_balance, dreamlayer_upload_image, dreamlayer_generate, dreamlayer_execution, dreamlayer_events and dreamlayer_download. Read the live tool schema before calling; do not pass fields the host does not expose.

Where this comes from

• Product overview: https://dreamlayer.io/agent?utm_source=clawhub&utm_medium=agent_distribution&utm_campaign=agent_launch_202609
• Canonical workflows and importer code: https://github.com/TheDesignFounder/dreamlayer-agent-plugin
• API reference and recovery: https://docs.dreamlayer.io/agent-api

Which job you are doing

Each job has a canonical workflow in the public repository TheDesignFounder/dreamlayer-agent-plugin under plugins/dreamlayer/skills/. Read the matching one and follow it rather than improvising:

All of them share one contract: docs/workflow-contract.md. It covers setup, the checks to run before spending, explicit operation selection, recovery and output acceptance.

The loop that actually works

1. Read dreamlayer_capabilities and dreamlayer_balance before paid work. Resolve any failure first. Confirm the operation you want is advertised and that the balance covers it.
2. Agree the brief, output count, aspect ratio, destination path and credit budget with the user. Never widen a batch or buy credits on your own.
3. Upload a reference once with dreamlayer_upload_image and keep its input_asset_id. Every operation except text_to_image needs one.
4. Save a unique idempotency_key before calling dreamlayer_generate, and name the operation explicitly: text_to_image, image_to_image, background_remove, upscale or sprite_sheet.
5. Work is asynchronous. dreamlayer_generate returns execution_id, status, asset, question and last_event_id, and a bounded stream often ends before the job does. Poll dreamlayer_execution, or resume from the cursor with dreamlayer_events. Never launch a replacement generation after a timeout.
6. A needs_input result is a question, not a failure. Answer it with respond and the conversation_id.
7. Download explicitly with dreamlayer_download to a new absolute path. A returned asset URL is not proof that a local file exists, and an existing file is never overwritten.
8. Inspect what you saved: dimensions, transparency when requested, spelling of any text, fidelity to the reference. Say so when an output has not been inspected.

Costs and failure states

Ordinary completed image operations currently cost one credit each; read the live figure from capabilities rather than assuming. Sprite jobs are a beta with frame-based pricing: read sprite_pricing, round the whole quote up once to a tenth of a credit, and pass it as max_credits. Transparent frames use the tier in first_tier_cents and additional_frame_cents; a sheet that keeps its background costs the flat plain_frame_cents per frame, roughly half, and is ordered with background: "keep" by a client that exposes the field. Sprite jobs cannot be cancelled by the customer; failed or expired jobs restore the hold.

Stop, keep the recoverable state, and report rather than looping paid retries when you hit: a missing or invalid key, an invalid request, an unsupported operation, insufficient credits, a failed or unfinished job, or a local download error. For a download error, repair the destination and download the same execution_id again.

What this does not do

• An API key is required, and paid generation spends DreamLayer API credits.
• Logo output is raster. There is no vector or SVG export, no trademark clearance and no legal clearance of any kind.
• Print artwork still needs a resolution, alpha and print-spec review before manufacture. There is no print certification.
• Nothing is published to any store, marketplace or ad platform for you.
• Product detail can drift during a generative edit, so review before publishing.
• Upscale doubles each side against a 4096 pixel ceiling, so an input whose longest side is over 2048 is refused before it costs a credit.
• Sprite output is a beta: broad requests do not guarantee correct motion, unseen detail or clean transparency.
• Godot import is validated end to end, including transparent frames, per-frame timing and a looping animation cycle. Unity editor validation is not confirmed; the Unity workflow ships importer code and instructions only.