AI Image Generator Skill

This entry points to a reviewed Beatra package for generating, composing, and editing product images, brand visuals, posters, social graphics, illustrations, and photo variations. The catalog entry contains no client code and performs no operation itself. Its review covers version 0.1.4, a pinned source tree, and the archive digest recorded in the entry; a newer package needs a fresh review. The package uses the MIT No Attribution (MIT-0) license.

Installation starts with explicit permission to download an external package into a temporary location. Verify the archive against the digest anchored in this entry, inspect its contents and scripts, and stop on a mismatch. The reviewed archive contains regular files without symlinks, executable bits, or binaries; verify that every file is byte-identical to the pinned source tree. Comparing the archive against the tree on the independent source host provides that check. Copying it into a new installation directory needs separate approval and must not overwrite an existing installation.

Automatic updates default to on. Disable them immediately after copying, before authorization, verification, tool discovery, uploading, or calling a tool:

python3 "$dest/scripts/mcp_client.py" update --auto off

This writes local updater state without making a network request. The setting depends on the resolved installation path and must be repeated after moving the package. Automatic updates otherwise check periodically and replace owned files using a CDN-provided checksum, which is not independent verification. Approve and inspect any later replacement separately.

Use the package only when the user explicitly chooses Beatra and approves paid rendering and third-party uploads. Do not use it for a local-only requirement or without spending approval. Read the installed skill and references for current routes, payloads, and the review loop. Free discovery can list current models and tasks before preparing a cost card.

Each cost card covers exactly one paid call: the tool, model, estimate, and files to upload. Upload only approved, user-provided files. Submit once with a stable client_request_id and poll the returned task. If the submission response is uncertain, retry the same identifier with identical arguments; do not replace a task that may still be running. Review the artifact with the user, report billing.net_charged_credits as the actual charge, and obtain fresh approval for every additional stage.

Estimates are not spending caps. Actual usage can exceed them and produce a negative prepaid balance. Provider-attributable failures receive refunds under the described policy; other charges are generally nonrefundable. Model availability and prices can change. Outputs are probabilistic and are not guaranteed unique or noninfringing, so review them before publication.

Inputs include an approved prompt, model choice, and any explicitly approved reference media. Outputs include task results, generated artifacts, and billing information. Users must have rights to uploaded material and avoid deceptive impersonation. The service transfers the output rights it owns, without guaranteeing third-party rights.

Browser device authorization requires separate permission. It grants a full-scope bearer token, including wallet:spend, with a sliding 15-day idle lifetime shared with other Beatra skills. Credentials use a restricted local file and directory; never expose the token in arguments, environment variables, or output.

The package requires Python 3.10+, a Beatra account, network access, and prepaid credits. It uses standard-library Python scripts without dependency installation or hooks. Authorization, hosted MCP operations, and package updates contact Beatra services. Prompts and media reach Beatra and the third-party model provider executing the task. Beatra is hosted outside mainland China.

Local state includes installation references, registered skill paths, updater state, and backups. Registration and calls send device or platform details and package identifiers without an opt-out. Account data is retained for the account lifetime; task, artifact, and billing records can remain after closure for audit or tax purposes.

Uninstallation begins with a dry run and inspection of its decision. Removing the last skill can revoke credentials and remove registrations; keeping a shared connection leaves its token in place. Updater state needs separate cleanup when appropriate. A failed revocation requires console revocation or expiry after the idle period.