claude-code-asset-generator

# Blueprint Studio

One plugin for Blueprint Studio Styles, assets, and durable image generation. Additional brand-context tools expose published guides and official assets when enabled on the connected server. Workflow skills and agents are optional; tools can be used directly.

This is the existing `blueprint-studio` plugin, upgraded in place. Its historical repository name remains `claude-code-asset-generator` to preserve installs. The public [Blueprint marketplace](https://github.com/Blueprint-Studio-AI/claude-code-marketplace) remains the discovery source. No second plugin or private brand snapshot is required.

## Install

Claude Code:

```text
/plugin marketplace add Blueprint-Studio-AI/claude-code-marketplace
/plugin install blueprint-studio@blueprint-studio-marketplace
```

For a local branch preview, start Claude Code with `--plugin-dir /absolute/path/to/this/repo`. Avoid adding a second standalone MCP connection if the plugin already supplies `asset-generator`.

Codex uses the portable package with `.codex-plugin/plugin.json` retained for older clients. Install the same package through the existing marketplace:

```text
codex plugin marketplace add https://github.com/Blueprint-Studio-AI/claude-code-marketplace.git
codex plugin add blueprint-studio@blueprint-studio-marketplace
```

Complete the host's MCP OAuth flow when prompted. Official ChatGPT/Codex and Claude directory reviews remain separate from direct installation.

Other skill-compatible hosts can use `skills/` and connect their remote MCP client to `https://tools.blueprintstudio.ai/api/mcp`. `plugin.json` and `mcp.json` provide the portable Agent Plugins manifest. Host transport spellings differ; compatibility manifests are intentionally retained.

Sign in through the host's MCP OAuth flow for account-wide access. Call `list_brands` to discover accessible organizations, then pass explicit `brandId` on every workspace call. Null/omitted selects personal scope, not the last-used brand. Legacy scoped credentials cannot switch workspaces. A plugin install grants no membership. If an existing connection only lists one workspace despite broader account access, renew its OAuth authorization and start a fresh client/thread. Older workspace-only grants are not silently expanded; a new API key is not the repair. Service API keys remain available for scoped automation; inference provider keys are not needed and credentials never belong in checked-in files.

## Tool contract

`generate_asset` already returns a durable `jobId` immediately. Use a fresh UUID `requestId` for each intended generation; reuse the same ID with identical arguments only for transport retries. Poll `get_generation_status` in the same `brandId` until `completed` or `failed`. Unknown outcomes and polling timeouts never establish failed generation or refunded credits. Preserve returned asset, receipt, and Style IDs. See [the MCP contract](skills/asset-generator/references/mcp-contract.md) for recovery and handoff details.

| Core tools | Context and references |
| --- | --- |
| Account OAuth and per-call workspace selection | `get_workspace_context` and published guide discovery |
| Styles, categories, generated examples, Style thumbnails | `list_models` capability and cost catalog |
| Durable generation/status, asset reads/downloads, background removal | `list_brand_assets` with version-pinned `generationInput` for ordered `inputs` |
| Brand/member administration within permissions | Exact `parentAssetId` edits and `get_generation_details` |

These capabilities require the current Blueprint server. Check live discovery for both tools and fields when connecting to an older deployment. Reference-set Style authoring, tracked forks, and Portal-backed tasks/approvals are not bundled features. There is no internal client list, private repository access, or automatic brand-site/database sync.

## Optional skills and agents

- `asset-generator`: browse, generate, inspect, refine, and deliver assets.
- `brand-manager`: requested workspace and Style administration.
- `style-gym`: optional repeatable Style experiments on diverse briefs.

Claude Code also discovers two agents from `agents/`: `blueprint-studio:asset-creator` for a bounded production brief and `blueprint-studio:style-evaluator` for requested Style comparisons. They inherit host tools and permissions, use the same `asset-generator` MCP connection, and add no mandatory routing. Other hosts can use the skills directly; agent auto-discovery is host-specific. See the [Claude Code plugin agent format](https://code.claude.com/docs/en/plugins-reference#agents).

Suggested handoff to an existing project agent:

> Use Blueprint Studio for this project's brand context and assets. Select the intended brand from list_brands and pass its brandId on each call. Inspect available Styles and any discovered brand guides/official assets. Continue the existing brief. Use tools directly or adapt the optional workflows; save selected asset/receipt/Style IDs and any unresolved job/request IDs with the work so another session can resume.

A project `.blueprint.json` may set local preferences. It is not an authorization mechanism. Do not duplicate live brand files into this plugin to personalize an installation.

## One source for every platform

Edit shared content here once. Do not maintain copies of skills or generation logic for each host.

| Authored source | Used by |
| --- | --- |
| `plugin.json` | Canonical name, release version, license, description, and OpenAI presentation |
| `mcp.json` | Canonical remote MCP connection |
| `skills/`, `agents/` | Shared workflows and optional host-specific agent discovery |
| Hosted Blueprint API | All tools, permissions, brand data, and generation behavior |

Run `python3 scripts/sync_distribution.py` after changing metadata. It generates `.claude-plugin/plugin.json`, `.codex-plugin/plugin.json`, `.mcp.json`, and `server.json`. Commit the generated files so hosts can install directly from Git. OpenAI and Cursor support the root Agent Plugins format; a separate Cursor source copy is unnecessary.

The historical marketplace repository is a stable discovery pointer, not another copy of the toolkit. Plugin releases take their version and metadata from this repository. See [distribution and release instructions](docs/distribution.md) for local checks and official review boundaries.

## Development

Backend and permission services live in the private monorepo; this public repository contains packaging and optional guidance only. Do not add tokens, production fixtures, client-private notes, or internal source links. Validate compatibility manifests and each skill before release. Public directory submission needs its own OAuth, metadata, and tool-annotation review.

Run the dependency-free checks locally:

```sh
python3 scripts/sync_distribution.py --check
python3 -m unittest discover -s scripts -p 'test_*.py'
claude plugin validate .
```

Also run the Codex plugin-creator `validate_plugin.py` and skill-creator `quick_validate.py` on each `skills/*` directory when available. Validation requires no generation calls or credentials. Installation, local cachebusting, and official directory release are distinct steps; bump the canonical release version rather than independently editing compatibility versions.

## License

This public plugin package is [MIT licensed](LICENSE). This does not license Blueprint Studio's separately hosted backend, private Style instructions, or client materials. Published MCP Registry metadata is separately dedicated to CC0 under the [Registry terms](https://modelcontextprotocol.io/registry/terms-of-service).