albumentationsx-mcp 0.6.0

Model Context Protocol server for AlbumentationsX augmentation discovery, validation, and previews.

AlbumentationsX MCP

Model Context Protocol server for AlbumentationsX: discovering transforms, validating augmentation pipelines, rendering deterministic previews, and exporting reproducible pipeline specs.

Scope

This project is intentionally a thin MCP layer around existing AlbumentationsX primitives:

• albu-spec is the source of transform metadata, parameter constraints, targets, and docstrings.
• albumentationsx remains the execution engine for validation, serialization, and previews.
• the MCP server exposes resources, tools, and prompts with strict typed schemas and bounded local file access.

The server does not execute arbitrary Python, fetch remote images, overwrite datasets, or train models.

Install

Run the published MCP server without cloning:

uvx --from albumentationsx-mcp albumentationsx-mcp

For local development:

uv sync --all-extras --dev

Run

uv run albumentationsx-mcp

Claude Desktop or another JSON-configured MCP host can launch a local checkout with stdio:

{
  "mcpServers": {
    "albumentationsx": {
      "command": "uv",
      "args": ["run", "albumentationsx-mcp"],
      "cwd": "/path/to/albu-mcp"
    }
  }
}

Installed from PyPI:

{
  "mcpServers": {
    "albumentationsx": {
      "command": "uvx",
      "args": ["--from", "albumentationsx-mcp", "albumentationsx-mcp"]
    }
  }
}

See examples/claude_desktop_pypi_config.json, examples/cursor_mcp_config.json, and examples/codex_mcp_config.toml for copyable host snippets.

Core Tools

• search_transforms: search transform metadata by query, targets, type, and bbox support.
• get_transform_schema: inspect a transform schema and constraints.
• validate_pipeline: validate a typed pipeline spec before running it.
• recommend_pipeline: create a conservative task preset for classification, detection, segmentation, or OCR.
• adjust_pipeline: apply structured preview feedback such as too_noisy or too_blurry.
• explain_pipeline: summarize likely effects, preview risks, and useful feedback tags.
• list_feedback_tags: list the structured feedback contract used by adjust_pipeline.
• list_quality_profiles: list task-aware quality profiles for balanced, classification, detection, segmentation, and OCR review.
• render_preview: create deterministic local preview artifacts inside an allowed output root.
• render_preview_batch: create deterministic multi-image preview contact sheets using the same request schema.
• compare_preview_runs: compare two preview manifests before choosing feedback tags or exporting a pipeline.
• summarize_tuning_session: summarize quality findings, feedback tags, score, risk, and export readiness.
• rank_preview_candidates: rank several candidate preview runs against one baseline.
• record_tuning_decision: persist a local accepted or rejected tuning decision.
• list_tuning_decisions: list local tuning decisions newest-first or score-ranked.
• export_tuning_report: export persisted tuning decisions as Markdown or JSON.
• list_preview_runs: list recent preview manifests recorded under the artifact root.
• get_preview_manifest: read one recorded preview manifest by run id.
• delete_preview_run: delete one preview run and its artifacts.
• cleanup_preview_runs: prune older preview runs beyond a retention count.
• export_pipeline: export a pipeline as Python, JSON, or YAML.

render_preview and render_preview_batch support optional bboxes, keypoints, and mask paths for annotation overlay previews. Preview manifests include an agent-legible summary block with input counts, seeds, transform names, artifact counts, contact sheets, and warnings.

What Changed In 0.2

• PyPI and MCP Registry publishing are automated with release version guardrails and post-release smoke checks.
• render_preview_batch produces batch previews and contact sheets for multi-image review.
• compare_preview_runs summarizes baseline and candidate manifests to compare preview runs before choosing feedback tags.
• Preview manifests include reproducibility summaries for seeds, transforms, artifact counts, and contact sheets.
• agent workflow resources and prompts guide preview tuning, annotation review, feedback adjustment, and final export.

What Changed In 0.3

• adjust_pipeline accepts optional feedback severity suffixes such as too_noisy:low, too_noisy:medium, and too_noisy:high.
• compare_preview_runs returns suggested_feedback_tags for candidate transforms that deserve visual review.
• Suggested tags are review candidates only; the user still chooses feedback after inspecting contact sheets.

What Changed In 0.4

• compare_preview_runs includes local quality_summary metrics for preview image artifacts.
• summarize_tuning_session explains baseline-to-candidate feedback, quality deltas, and export readiness.
• task workflow profiles and docs/RECIPES.md guide classification, detection, segmentation, and OCR MCP host sessions.

What Changed In 0.5

• quality_summary now includes saturation, colorfulness, entropy, clipping, and deterministic quality findings.
• Annotation previews record bbox, keypoint, and mask retention observations in preview manifests.
• compare_preview_runs includes annotation_summary when annotation observations are available.
• summarize_tuning_session returns quality_score, quality_risk, and structured quality_findings.
• record_tuning_decision and list_tuning_decisions provide a local JSON-backed tuning decision journal.

What Changed In 0.6

• Added task-aware quality profiles for balanced, classification, detection, segmentation, and OCR review.
• Added rank_preview_candidates to choose between multiple candidate preview runs.
• Added export_tuning_report for Markdown or JSON handoff from the local tuning decision journal.
• Extended golden MCP evals to cover two-candidate ranking and report export.

Demo Workflow

1. Use recommend_pipeline and validate_pipeline for a conservative baseline.
2. Call render_preview_batch on a small local image set.
3. Adjust with structured feedback such as too_noisy, too_noisy:high, or too_distorted.
4. Render the candidate preview batch with the same inputs.
5. Call compare_preview_runs before accepting the candidate and inspect quality_summary.findings.
6. If several candidates exist, call rank_preview_candidates with the matching quality profile.
7. Call summarize_tuning_session and review quality_score, quality_risk, and export_ready.
8. Call record_tuning_decision for the accepted or rejected candidate.
9. Call export_tuning_report for handoff, then export the accepted pipeline with export_pipeline.

See docs/USAGE.md for an end-to-end MCP host workflow, docs/RECIPES.md for task-specific host recipes, docs/DEMO.md for a generated preview comparison demo, CHANGELOG.md for release notes, docs/RELEASE.md for the package and MCP Registry release process, server.json for public discovery metadata, and evals/golden_mcp_scenarios.yaml for executable MCP scenarios.

Verification

uv run pytest
uv run ruff check .
uv run ruff format --check .
uv run ty check
uv run python scripts/run_golden_evals.py
uv build