# AlbumentationsX MCP

Model Context Protocol server for [AlbumentationsX](https://github.com/albumentations-team/AlbumentationsX): inspect datasets, preview augmentations, refine them with visual feedback, and export reproducible pipelines.

<!-- mcp-name: io.github.dKosarevsky/albu-mcp -->

[![CI](https://github.com/dKosarevsky/albu-mcp/actions/workflows/ci.yml/badge.svg)](https://github.com/dKosarevsky/albu-mcp/actions/workflows/ci.yml)
[![PyPI](https://img.shields.io/pypi/v/albumentationsx-mcp)](https://pypi.org/project/albumentationsx-mcp/)
[![Python](https://img.shields.io/badge/python-3.10--3.13-blue)](pyproject.toml)
[![MCP Registry](https://img.shields.io/badge/MCP%20Registry-active-green)](https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.dKosarevsky/albu-mcp)
[![skills.sh](https://skills.sh/b/dKosarevsky/albu-mcp)](https://skills.sh/dKosarevsky/albu-mcp)

![Baseline and adjusted AlbumentationsX preview contact sheets](https://raw.githubusercontent.com/dKosarevsky/albu-mcp/main/docs/assets/demo/comparison_contact_sheet.png)

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`](https://github.com/dKosarevsky/albu-mcp/releases/latest/download/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:

```bash
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](docs/CONFIGURATION.md). Copyable host configurations are in [the install guide](docs/INSTALL.md).
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:

```text
Run the host smoke check. If preview_ready is true, call run_first_preview for /absolute/path/to/images with low
intensity and at most 8 images. Show me the contact sheet. When I mention a specific result, call
trace_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](docs/use-cases/CLASSIFICATION_ROBUSTNESS.md), or follow the
[First 10 Minutes guide](docs/FIRST_10_MINUTES.md). The `validate_preview_request` fallback, batch previews, and how to
compare preview runs are in [Usage](docs/USAGE.md). Use `too_noisy:high` or `exposure_too_weak:medium`, then optionally
share one redacted loop through [first-preview feedback](docs/FIRST_PREVIEW_FEEDBACK.md).
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](https://albumentations.ai/docs/integrations/mcp/)
- [Official MCP Registry entry](https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.dKosarevsky/albu-mcp)
- [skills.sh agent skill](https://skills.sh/dKosarevsky/albu-mcp)
- [Upstream documentation PR](https://github.com/albumentations-team/AlbumentationsX/pull/289)

## Documentation

- [Install and host configuration](docs/INSTALL.md)
- [Runtime settings and capability profiles](docs/CONFIGURATION.md)
- [First 10 minutes](docs/FIRST_10_MINUTES.md)
- [First-preview feedback](docs/FIRST_PREVIEW_FEEDBACK.md)
- [Usage](docs/USAGE.md) and [recipes](docs/RECIPES.md)
- [CPU Tensor Compose validation and export](docs/TORCH_CPU_COMPOSE.md)
- [MCP Apps review](docs/MCP_APPS_REVIEW.md) and [compatibility policy](docs/COMPATIBILITY.md)
- [Documentation index](docs/INDEX.md)
- [CHANGELOG.md](CHANGELOG.md)
- [server.json](server.json): public MCP Registry metadata.

## Development

```bash
uv sync --all-extras --dev
uv run pytest
uv run ruff check .
uv run ruff format --check .
uv run ty check
```

Licensed under [AGPL-3.0-or-later](LICENSE).