Is your unknown MS/MS spectrum really that library compound?

Paste the query spectrum and the candidate library spectra. Your browser scores every pair with cosine and modified cosine, checks precursor, ion mode and adduct, and draws the mirror plot - free, nothing uploaded. A paid run then annotates the feature with a confidence level or writes the report text.

Each example has a saved model run, so you can see the whole page for free.

Feature

Query spectrum (MGF, MSP, MassBank, JSON or "m/z intensity" lines; a file of many features lets you pick one)

Drop a file or

Candidate library spectra (MGF, MSP, MassBank or JSON, up to 2,000)

Drop a file or

Query overrides and scoring settings

Query precursor m/z

Query ion mode

Query adduct

Fragment tolerance (Da)

Minimum score

Minimum matched peaks

Relative-intensity floor

Precursor tolerance (ppm)

Candidates sent to a run

There is no universal identification threshold: scores depend on processing, mass accuracy, collision energy and library quality. Defaults follow the source skill's library-search example (score 0.6, 5 matched peaks, 0.02 Da).

Notes (optional: sample, LC, instrument, collision energy, library source, and any standard, isotope or retention-time evidence)

Load set .json

Scores (free, in your browser, no AI)

Ref	Candidate	Cosine / matched	Modified / matched	Precursor delta Da (ppm)	Explained	Status

Mirror plot

Query above, reference below. Matched peaks in the accent colour, precursor-shifted matches in amber. Click a candidate's name to plot it.

Flags

Free: no account, no run.

Anything to ask (optional)

Paste a query and a library to price the run.

Raw model reply

The reply did not parse as the structured result, so it is shown as it arrived.

Your recent runs

What this does, and what it does not

The scores follow matchms exactly: peaks within the tolerance are paired, pairs are taken greedily by intensity product so each peak is used once, and the sum is divided by both spectra's norms (m/z power 0, intensity power 1). The modified cosine also pairs peaks shifted by the precursor difference, which is how analogs - a methyl group more or less - show up. The implementation was checked against the matchms 0.33.1 source on 3,000 scores with no difference beyond floating-point rounding. Query and references are processed the same way, as the source skill insists: invalid peaks dropped, intensities scaled to the base peak, peaks under the relative-intensity floor removed.

The paid run assigns a confidence level on the scheme of Schymanski et al. (Environ. Sci. Technol. 2014): level 1 needs an authentic standard measured under the same conditions, level 2a a library spectrum match, level 3 a tentative candidate. A library match alone never reaches level 1, and the model is told never to rescore spectra or invent a fragment. Derived from the agent skill @k-dense-ai/matchms (k-dense-ai/scientific-agent-skills, K-Dense Inc.). The example spectra are illustrative, not measured.