FIT Kotlin Multiplatform SDK

Reads and writes Garmin FIT files. Profile 21.213.0-Release.

Everything under src/ is machine-generated, with one exception: src/jsMain/resources, the browser demo, is hand-written and local to this repository.

The two distributions carry the same profile version, which is the SDK's version: the badges above are the latest of each.

// Gradle Kotlin DSL
dependencies {
    implementation("io.github.glandais:fit-kotlin-sdk:21.213.0")
}

KMP consumers resolve their own variant through Gradle module metadata; the published per-target artifacts are -jvm, -js, -wasm-js and -wasm-wasi. A plain JVM (or Java) project can depend on fit-kotlin-sdk-jvm directly.

npm install @glandais/fit-kotlin-sdk

Ships .d.ts definitions and runs in both browsers and Node.js. The API it exposes is not the Kotlin one — see JavaScript and TypeScript.

• Gradle 9.x with JDK 25, or Gradle 8.x with JDK 17 or 21. Gradle 8.14 does not run on JDK 25 at all, so on that combination set JAVA_HOME=/path/to/jdk21.
• Targets: jvm, js and wasmJs (both node and browser), and wasmWasi. The SDK itself is all in commonMain with no expect/actual and no java.* APIs, so adding a native target to build.gradle.kts needs no source changes. jsMain holds the JavaScript/TypeScript surface only — see JavaScript and TypeScript.

On the js target — not the Wasm ones, which have real types — Byte, Short, Int, Float and Double are the same runtime type, a JS number. Nothing in the API depends on telling them apart except [BaseType.of], which infers the width of a field the profile does not know from the first value set on it: there it answers FLOAT64 for any number, holding the value rather than truncating it. Set such a field through a Field you built yourself if you need an exact width on that target.

decode() never throws. A truncated or corrupt file still yields everything that decoded before the problem, alongside a description of it:

val result = FitDecoder(bytes).decode()

result.errors.forEach { println("${it.message} at byte ${it.bytePosition}") }

result.messages.recordMesgs.forEach { record ->
    println("${record.timestamp}  ${record.heartRate} bpm  ${record.distance} m")
}
result.messages.sessionMesgs.forEach { session ->
    println("${session.sport}  ${session.totalElapsedTime} s")
}

Values are typed by the profile: scaled fields arrive in their real units as Double, enums as generated enum classes, timestamps as kotlin.time.Instant. Everything nullable, because FIT has no null and an absent field is indistinguishable from one holding the invalid sentinel. Use Field.getRawValue when that difference matters.

For files too large to hold decoded. These do throw FitFormatException, having no result object to carry an error in:

FitDecoder(bytes).asSequence()
    .filterIsInstance<RecordMesg>()
    .forEach { println(it.heartRate) }

FitDecoder(bytes).read { mesg -> println(mesg) }

Both, and decode(), are the same pull-based cursor underneath.

The SDK has no dependencies, so it ships no Flow entry point. Decoding is synchronous and CPU-bound, so one is a one-liner over the sequence:

fun FitDecoder.asFlow(options: DecodeOptions = DecodeOptions()): Flow<Mesg> =
    asSequence(options).asFlow().flowOn(Dispatchers.Default)

FitDecoder(bytes).decode(
    DecodeOptions(
        mode = DecodeMode.NORMAL,     // or SKIP_HEADER, DATA_ONLY
        expandComponents = true,      // unpack bit-packed fields
        includeUnknownData = false,   // keep messages absent from the profile
        mergeHeartRates = true,       // fold `hr` messages into `record`s
    ),
)

The typed classes are a view over numbered fields; nothing is hidden behind them:

val mesg: Mesg = result.mesgs.first()
mesg.getFieldValue(RecordMesg.HEART_RATE_FIELD_NUM)
mesg.fieldList.forEach { println("${it.fieldName} = ${it.toList()} ${it.units}") }

A FIT file must open with a file_id message. close() is not optional: it back-patches the header's data size and appends the file CRC. encodeFit {} calls it for you.

val bytes = encodeFit {
    write(
        FileIdMesg().apply {
            type = File.ACTIVITY
            manufacturer = Manufacturer.GARMIN
            timeCreated = Clock.System.now()
        },
    )
    write(RecordMesg().apply { heartRate = 140u; distance = 1234.5 })
}

Message definitions are emitted only when the layout changes, so a run of same-shaped records costs one definition rather than one each.

Never redefine an existing profile message or field: that breaks every other reader. FIT's sanctioned extension mechanism is developer fields, declared in the file itself by a developer_data_id / field_description pair. The decoder surfaces them on Mesg.developerFieldList, and the declarations it met on FitMessages.developerFieldDescriptions.

To write them, declare each field with FitEncoder.registerDeveloperField before the first message carrying it — the encoder emits the declaring message pair for you. DeveloperFieldDescription.createField() then makes a field to setDeveloperField on a message; re-registering the descriptions a decode returned re-encodes the same extension fields.

The js target carries one source set the other targets do not, jsMain, whose whole job is to turn the Kotlin API into an idiomatic JavaScript one: plain objects instead of Kotlin classes, Dates instead of FIT timestamps, "cycling" instead of 2. It is published to npm as @glandais/fit-kotlin-sdk with TypeScript definitions, and the exported names are flat — the source set has no package, precisely so that consumers write decodeFit and not com.garmin.fit.decodeFit. Install it with npm.

import { decodeFit, isFitFile, fitFieldInfo, fitProfileVersion } from '@glandais/fit-kotlin-sdk'

const bytes = new Int8Array(await file.arrayBuffer())
if (!isFitFile(bytes)) throw new Error('not a FIT file')

const result = decodeFit(bytes)                    // never throws; see result.errors
for (const record of result.messages.recordMesgs ?? []) {
    console.log(record.fields.timestamp, record.fields.heartRate, record.fields.positionLat)
}

decodeFit returns

{
    profileVersion: number | null,
    errors: { message: string, bytePosition: number }[],
    messages: Record<string, FitMessage[]>,   // { recordMesgs: [...], sessionMesgs: [...] }
    mesgs: FitMessage[],                      // the same objects, in file order
}

// FitMessage: { name, num, index, fields: Record<string, any>, developerFields: Record<string, any> }

which is the shape @garmin/fitsdk produces, so code written against the official JavaScript SDK ports across with little more than a rename. Options, all optional:

decodeFit(bytes, {
    mode: 'normal',            // or 'skipHeader', 'dataOnly'
    expandComponents: true,    // unpack bit-packed fields
    includeUnknownData: false, // keep messages absent from the profile
    mergeHeartRates: true,     // fold `hr` messages into `record`s
    applyTypes: true,          // enum values as names, date_time as Date, bool as boolean
})

applyTypes is the one knob with no Kotlin equivalent: the Kotlin API returns Sport.CYCLING from a generated accessor, and JavaScript has no such accessors, so the profile's names are resolved during the decode instead. Turn it off to see exactly what the file stores.

Writing works the same way round — names or numbers, Dates or FIT seconds:

import { encodeFit } from '@glandais/fit-kotlin-sdk'

const bytes = encodeFit([
    { name: 'fileId', fields: { type: 'activity', manufacturer: 'garmin', timeCreated: new Date() } },
    { name: 'record', fields: { timestamp: new Date(), heartRate: 140, cadence: 90 } },
])

fitFieldInfo('record', 'heartRate') reports what the profile says about a field — number, units, scale, offset, base type, profile type — which is enough to label a table without a lookup table of your own.

The package is a UMD bundle, which named imports above reach through any bundler (webpack, Vite, esbuild) and through TypeScript. Node's own ESM loader is stricter and needs the default import:

import pkg from '@glandais/fit-kotlin-sdk'
const { decodeFit } = pkg
// or: const { decodeFit } = require('@glandais/fit-kotlin-sdk')

The npm version is the Maven Central one, unchanged: this SDK releases under three numeric components only, major.minor from the FIT profile and a patch that increments for a release at an unchanged profile.

One limit worth stating: a uint64 beyond 2^53 arrives rounded, because a JavaScript number has 53 bits of mantissa. Only the FIT *_64 types and a few device serial fields reach that far.

src/jsMain/resources holds a decoder demo — drop a .fit file in, get a summary, the track, the series, and every message in a table. Unlike src/, it is not generated, so it is the one place in this repository that can be edited by hand.

gradle demo          # -> build/dist/js/productionExecutable

Serve that directory over HTTP (python3 -m http.server) — file:// will not do, the sample activity is fetched. It is published to GitHub Pages on every push to develop by .github/workflows/gh-pages.yml: https://glandais.github.io/fit-kotlin-sdk/.

gradle build
gradle jvmTest      # 198 tests on the JVM
gradle jsNodeTest   # the same tests plus the JS surface, on Node
gradle demo         # the browser demo, into build/dist/js/productionExecutable

• Protocol and profile docs: https://developer.garmin.com/fit
• Cookbook recipes: https://developer.garmin.com/fit/cookbook/