plantcv-mcp

plantcv-mcp

plantcv-mcp is an MCP server that lets AI agents run PlantCV plant phenotyping. It returns trait numbers and the segmentation overlay they were measured from, and refuses to return numbers when the segmentation is degenerate.

[CI] [PyPI] [Python] [License] [DOI] [Glama]

[Correct segmentation: four views of a rendered plant, stems and leaves tinted red and outlined in cyan, the gray background untouched] [Inverted segmentation: the whole gray background tinted red, the four plant views left out]

Left, channel="a", object_type="dark": the mask covers 3.1% of the frame, 9 components, area=32427. Right, channel="s", object_type="dark": the mask covers 96.1%; it is the background, area=1007829.

Why you get the overlay

Both images come from the same file and the same threshold method; only the channel differs. Without the picture, both runs return 17 traits with correct units and believable magnitudes. The one on the right is measuring the wall behind the plants. Red marks the measured pixels; a cyan line traces the mask's edge.

So segment() returns the overlay and diagnostics but no traits, and measure() needs the session_id that segment() mints: no single image is measured before you are shown it. The exception is measure_images(), which applies one recipe to up to 200 images without overlays. Settle the recipe on one image's overlay first (the server does not check); every image still passes the segment() guards, and one that trips a blocking guard gets no traits.

Measured with PlantCV 4.11.3 on a synthetic render (first row: tests/fixtures/multi_specimen.png) and on photographs, each of these returns believable numbers with correct units:

Install

With uv nothing needs installing: uvx --python 3.13 plantcv-mcp fetches and runs the current release. Otherwise:

pip install plantcv-mcp

Requires Python 3.11 to 3.13: PlantCV 4.11.3 needs SciPy below 1.16, which has no Python 3.14 wheels, so --python 3.13 keeps uvx off a 3.14 default. PlantCV and its scientific stack make the first install (or first uvx run) slow. From a checkout: uv add /path/to/plantcv-mcp. The one optional tool, segment_leaves_sam(), needs pip install "plantcv-mcp[sam]" (torch, torchvision, segment-anything); without it, that tool refuses with that command and every other tool works.

Configure your MCP client

claude mcp add plantcv -- uvx --python 3.13 plantcv-mcp

Claude Desktop and other stdio hosts:

{ "mcpServers": { "plantcv": { "command": "uvx", "args": ["--python", "3.13", "plantcv-mcp"] } } }

With a pip install the command is plantcv-mcp (checkout config). Verify with list_methods().

Flags: --root DIR (repeatable, or PLANTCV_MCP_ROOTS) confines every read, and every file the server keeps, to your imagery: plantcv-mcp --root /data/phenotyping. --no-isolate (or PLANTCV_MCP_ISOLATE=0) runs analyses in-process instead of in the crash-containing worker.

What the benchmark shows

Claude with the server was compared with Claude writing its own code (mostly OpenCV) in a Python environment we set up with the same PlantCV and OpenCV, on 46 measurement questions about rendered plant images, 3 runs each. A tie means the server matches that result without you building the environment. Questions right in every run, server against Claude's own code: Claude Haiku 5.5 38 against 32, Sonnet 5.5 33 against 28, Opus 5.5 42 against 44. None is separated (pre-registered; exact McNemar p = 0.15, 0.23 and 0.63), and Claude's own code was cheaper on all three models.

[Stacked bar chart of right and wrong answers in the 2026-10-18 re-run, 46 questions times 3 runs per setup. Claude + plantcv-mcp + Read: 132 right, 6 wrong. Claude + Bash/PlantCV/OpenCV + web: 134 right, 4 wrong. No-model controls, one answer per question: one color cut 28 right, 10 wrong, 8 declined; median answer of its kind 2 right, 44 wrong.]

The chart is the Opus 5.5 re-run on 2026-10-18, against an unreleased commit (051aff0) that adds several-channel segmentation rules and zoom_image: right on 97.1% of answers with its own code and 95.7% with the server. Its own code cost $0.038 against $0.066 per answer (medians; the intervals do not overlap). On release 1.16.1 the server was right on 86.2% against 95.7% (p = 0.18). The images have crisp edges, so a single color cut with no model answers 28 of the 46.

We wrote the questions and ran the benchmark ourselves; an outside re-run is welcome. Method, controls and limitations: benchmarks/agent_qa; the Haiku and Sonnet runs: 2026-10-20 re-run.

Tools

Typical loop: suggest_segmentation → segment → look at the overlay → segment again with another channel, method or polarity if it is wrong (or refine if it is nearly right) → measure. With a ColorChecker in the frame, color_correct=true corrects colors before segmenting and excludes the card from the mask. Parameters, an example response and the measured facts behind each tool: docs/GUIDE.md.

What it refuses

Each guard was added after a specific failure and names the next action. Blocking guards withhold numbers; advisories travel with them.

• Inverted mask (implausible_coverage): the right-hand image above, 17 believable traits describing the wall.
• Nothing selected, or fill_size deleted the specimen (empty_mask, fill_erased_mask): PlantCV returns 17 zeros with in_bounds=True.
• Background texture (noisy_segmentation): specks of chamber wall measured as one plant.
• Several plants in one mask (multi_specimen): use measure_regions().
• Wrong scale, wrong session kind, a file changed after segmentation: refused, naming the right tool.
• A lens calibration the boards do not determine, or no color card when one was asked for: refused.

Why each guard exists, and every warning code: docs/GUIDE.md#warnings-and-refusals.

Security

With no --root, the server reads any image file the host user can read and returns it to the model. It writes only the lens-corrected image, the Segment Anything checkpoint when asked to download it, and short-lived temporary copies of hyperspectral and thermal inputs. Run it as a non-root user with --root set. Analyses run in a worker subprocess, so a native crash is a tool error, not a dead server. Details: docs/GUIDE.md#security-and-trust-boundary.

Attribution and license

MIT licensed. It depends on PlantCV, which is licensed under the Mozilla Public License 2.0. No PlantCV source is vendored or redistributed here; it is an ordinary runtime dependency, so the MIT license applies to this project's own files. Unofficial: not affiliated with, endorsed by, or sponsored by the Donald Danforth Plant Science Center or the PlantCV maintainers. See NOTICE.

Cite it with CITATION.cff (archived at doi:10.5281/zenodo.21713516). Changes: CHANGELOG.md.