AlbumentationsX MCP

AlbumentationsX MCP

Model Context Protocol server for AlbumentationsX: inspect datasets, preview augmentations, refine them with visual feedback, and export reproducible pipelines.

[CI] [PyPI] [Python] [MCP Registry] [skills.sh]

[Baseline and adjusted AlbumentationsX preview contact sheets]

Ask an MCP host for several robustness variants, reject an excessive result such as too_noisy:high, compare the adjusted batch previews, and export the accepted pipeline.

Install

Claude Desktop

Download the latest albumentationsx-mcp.mcpb, install it from Settings -> Extensions -> Advanced settings, and select separate image and artifact directories.

Other MCP Hosts

Run the published server with bounded local access:

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

run_first_preview requires the default full or dataset capability profile. The smaller review profile uses the explicit validate/render fallback in the usage guide, or you can restart with dataset or full; see configuration. Copyable host configurations are in the install guide. The repository also contains a native Codex plugin bundle. npx skills add dKosarevsky/albu-mcp installs agent guidance, not the MCP server.

First Preview

After connecting the server, ask your host:

Run the host smoke check. If preview_ready is true, call run_first_preview for /absolute/path/to/images with lowintensity and at most 8 images. Show me the contact sheet. When I mention a specific result, calltrace_preview_variant before adjusting it.

run_host_smoke_check returns preview_ready and a preview_request_template. If resource reads are unavailable, call get_workflow_example with example_id="client-smoke".

Try the classification robustness use case, or follow the First 10 Minutes guide. The validate_preview_request fallback, batch previews, and how to compare preview runs are in Usage. Use too_noisy:high or exposure_too_weak:medium, then optionally share one redacted loop through first-preview feedback. If setup fails, read albumentationsx://diagnostics/guide and call diagnose_environment for bounded remediation actions.

Capabilities

• Transform discovery, schemas, recipes, and pipeline validation.
• Classification, detection, segmentation, OCR, bbox, mask, keypoint, and dataset-quality workflows.
• Deterministic previews, contact sheets, annotation overlays, comparison, ranking, and reports.
• Interactive MCP Apps review with a text-only fallback for other hosts.
• Structured feedback, tuning sessions, and Python, JSON, or YAML export.
• Runtime-aware CPU torch.Tensor pipeline validation and guarded Python handoff.
• MCP 2026-07-28 plus legacy negotiation; stable agent workflow resources, diagnostics, and contract snapshots.

The server does not execute arbitrary Python, fetch remote images, overwrite datasets, or train models. Reads are restricted by --allowed-root; generated files stay under --artifact-root.

Integrations

• Official Albumentations MCP guide
• Official MCP Registry entry
• skills.sh agent skill
• Upstream documentation PR

Documentation

• Install and host configuration
• Runtime settings and capability profiles
• First 10 minutes
• First-preview feedback
• Usage and recipes
• CPU Tensor Compose validation and export
• MCP Apps review and compatibility policy
• Documentation index
• CHANGELOG.md
• server.json: public MCP Registry metadata.

Development

uv sync --all-extras --devuv run pytestuv run ruff check .uv run ruff format --check .uv run ty check

Licensed under AGPL-3.0-or-later.