Communitygithub.com

dKosarevsky/albu-mcp

MCP server for AlbumentationsX image augmentation workflows

¿Qué es albu-mcp?

albu-mcp is a Codex agent skill that mCP server for AlbumentationsX image augmentation workflows.

Compatible con~Claude CodeCodex CLI~Cursor
npx skills add dKosarevsky/albu-mcp

Preguntar en tu IA favorita

Abre un nuevo chat con esta habilidad de agente ya precargada.

Documentación

AlbumentationsX MCP

Install

uvx --from albumentationsx-mcp albumentationsx-mcp
uvx --from albumentationsx-mcp albumentationsx-mcp \
  --allowed-root /absolute/path/to/images \
  --artifact-root /absolute/path/to/albu-artifacts

First Run Prompt

Use this as the first host task:

Use AlbumentationsX MCP on image or directory `DATASET_PATH`. Read from `ALLOWED_ROOT` and write to `ARTIFACT_ROOT`. When the host exposes resource reads, read `albumentationsx://examples/client-smoke`; if resource reads are unavailable, call `get_workflow_example` with `example_id="client-smoke"`. Then call `run_host_smoke_check`. Continue only when `preview_ready` is true. If `preview_ready` is false, call `diagnose_environment` and stop before rendering. Then call `plan_dataset_onboarding`, `build_review_packet`, `validate_preview_request`, and `render_preview_batch`. Render at most 6 images on the first pass. Show the contact sheet path and ask for concrete feedback before `adjust_pipeline` or `export_pipeline`.

Host Config Hints

  • Codex plugin mode uses .codex-plugin/plugin.json and .mcp.json; its pinned server grants no user dataset root.
  • Set ALBU_MCP_ALLOWED_ROOTS and ALBU_MCP_ARTIFACT_ROOT, or use explicit absolute host args. Never rely on the working directory.
  • Restart, run run_host_smoke_check, and stop unless allowed_roots contains the intended root and preview_ready is true.

Host Workflow

  1. Read albumentationsx://examples/client-smoke; if resource reads are unavailable, call get_workflow_example with example_id="client-smoke".
  2. Call run_host_smoke_check next; continue only when preview_ready is true.
  3. Call plan_dataset_onboarding, then build_review_packet for one image or folder.
  4. Validate user paths with validate_preview_request before rendering.
  5. Render a small sample with render_preview_batch; inspect the contact sheet.
  6. Record feedback with record_preview_feedback, such as too_noisy:high or exposure_too_weak.
  7. adjust_pipeline, re-render, and compare before acceptance.
  8. Export only reviewed work with export_pipeline or requested report tools.

Safety Rules

  • Do not train, overwrite datasets, or fetch remote images.
  • Keep private paths, filenames, and image contents out of public reports.
  • Generated fixtures, contact sheets, and rehearsals are not beta evidence.
  • Read only under --allowed-root; write only under --artifact-root.
  • Re-run validation after changing paths, masks, bboxes, labels, or annotation formats.

Stop Conditions

  • Missing real image or dataset-directory path: ask for one.
  • Path outside --allowed-root: refuse that path and ask for a bounded path.
  • User asks for many variants: render a small first batch before expanding.

Evidence Workflows

For evidence work, follow the generated pack README and use:

  • albu-mcp activation real-adoption-cycle
  • albu-mcp activation product-fix-closure-pipeline
  • albu-mcp evidence execution-pack --date YYYY-MM-DD --reviewer "Release operator" --output-dir evidence-session --format markdown
  • albu-mcp evidence execution-pack-audit --input-dir evidence-session
  • albu-mcp evidence execution-pack-progress --input-dir evidence-session
  • albu-mcp evidence execution-pack-status --input-dir evidence-session --format markdown --output evidence-session/status.md
  • albu-mcp evidence preflight
  • albu-mcp evidence import-wizard

If preview setup fails, read albumentationsx://diagnostics/guide, call diagnose_environment, then remediate before rendering.

Skills relacionados