MNE MCP Guard

Use this skill to make MNE MCP execution reliable, especially before expensive steps (ICA, time-frequency, source localization) and on unfamiliar data.

Workflow

1. Check capability first — mne_check_status. If MNE is missing, stop and say so; if

scikit-learn is missing, ICA is unavailable.

1. Inspect before processing — mne_get_info / mne_describe. Never guess channel names,

sampling rate, montage, or event codes.

1. Look before parameterizing — mne_plot_psd (read the PNG) before choosing filter cutoffs;

mne_plot_raw before choosing rejection thresholds.

1. Smoke-test heavy steps

• Before ICA: confirm data is high-pass filtered (~1 Hz) and montage is set.
• Before epoching: verify real event codes exist (mne_find_events /

mne_events_from_annotations) and that the epoch window fits the recording.

• Before TFR: confirm the epoch window is long enough for the lowest frequency.

1. Escalate gradually — get a minimal version working (one condition, default params), then add

options. Don't jump to a 60-component ICA or full source pipeline on the first try.

1. Read warnings, not just success — dropped epochs, rank deficiency, montage subset warnings,

and annotation notes change the validity of results.

Guardrails

• Units are SI (volts/tesla). A reject of 100 is 100 volts. 100 µV is 100e-6. This is the

most common silent error — verify any threshold's order of magnitude.

• Set a montage before topomaps, ICA component plots, or interpolation.
• High-pass before ICA (~1 Hz) or components will be unstable.
• Don't assume a file loads in one call — BrainVision/EEGLAB need sidecar files; point at the header.
• Don't conclude "timeout = broken." ICA/TFR/large files are genuinely slow; raise

MNE_MCP_TIMEOUT and retry, or crop/decimate first.

Decision tree

1. "no session object named ..."? → mne_session_info; load the data or fix the name.
2. Plot/interpolation error mentioning positions/montage? → mne_set_montage; check channel

names match the montage.

1. ICA error or weird components? → confirm sklearn present, high-pass applied, and

n_components ≤ data rank (lower it or use a 0.99 variance fraction).

1. TFR "wavelet longer than signal"? → wider epochs, higher fmin, or smaller n_cycles via

mne_run_code.

1. All epochs dropped / empty evoked? → loosen reject_eeg, verify event codes, check the epoch

window against recording length.

1. Step times out? → increase MNE_MCP_TIMEOUT; for large raw use preload=false + mne_crop.

References

• Read references/failure-patterns.md for concrete error→fix mappings.
• Use mne-analyst for the full workflow and pipeline conventions; this skill is specifically for

failure prevention and recovery.