diff --git a/kotlin/build.gradle.kts b/kotlin/build.gradle.kts
index 40ca0be..dc4eadc 100644
--- a/kotlin/build.gradle.kts
+++ b/kotlin/build.gradle.kts
@@ -1,6 +1,5 @@
 plugins {
     kotlin("multiplatform") version "2.3.21"
-    id("com.squareup.wire") version "6.4.0"
     id("com.vanniktech.maven.publish") version "0.36.0"
 }
 
@@ -9,19 +8,24 @@ version = providers.gradleProperty("VERSION_NAME").getOrElse("0.2.3")
 
 repositories {
     mavenCentral()
+    maven("https://central.sonatype.com/repository/maven-snapshots/") {
+        mavenContent { snapshotsOnly() }
+    }
 }
 
 kotlin {
-    jvmToolchain(17)
+    jvmToolchain(21)
 
     jvm()
-    iosArm64()
-    iosSimulatorArm64()
 
     sourceSets {
         val commonMain by getting {
             dependencies {
-                api("com.squareup.wire:wire-runtime:6.4.0")
+                // Proto types (TAKPacketV2, GeoChat, etc.) come from the published
+                // protobufs SDK. Using `implementation` ensures we do NOT re-export
+                // them to consumers — they bring their own protobufs SDK dependency
+                // and there is exactly one source of truth on the classpath.
+                implementation("org.meshtastic:protobufs:2.7.25-SNAPSHOT")
             }
         }
         val jvmMain by getting {
@@ -41,91 +45,10 @@ kotlin {
     }
 }
 
-// Wire 6.2 KMP config — mirrors Meshtastic-Android's core:proto module so the
-// SDK's generated TAKPacketV2 type matches theirs exactly. The `protobufs`
-// submodule is the single source of truth; Wire reads `meshtastic/atak.proto`
-// directly from it and emits Kotlin to `build/generated/source/wire/` in the
-// commonMain source set.
-//
-// NOTE: the submodule contains ~22 other .proto files (mesh/config/admin/…)
-// and a top-level `nanopb.proto` that imports `google/protobuf/descriptor.proto`
-// — none of which we want Wire to try to compile. The `include(...)` filter on
-// sourcePath restricts Wire to just `atak.proto`, which is self-contained and
-// has no imports, so this keeps the build hermetic.
-wire {
-    sourcePath {
-        srcDir("../protobufs")
-        include("meshtastic/atak.proto")
-    }
-    kotlin {
-        // Skip defensive copies of repeated / map fields on decode — matches
-        // Meshtastic-Android's performance tuning.
-        makeImmutableCopies = false
-        // Sentinel value: flatten every oneof to nullable properties on the
-        // parent class. With 11 cases in TAKPacketV2.payload_variant, this
-        // produces `pli: Boolean?`, `chat: GeoChat?`, `casevac: CasevacReport?`,
-        // etc. as top-level fields instead of an intermediate sealed class.
-        boxOneOfsMinSize = 5000
-    }
-}
-
 tasks.withType<Test>().configureEach {
     useJUnitPlatform()
 }
 
-// ─────────────────────────────────────────────────────────────────────────────
-// JVM JAR packaging — proto classes are SHIPPED with two surgical exceptions.
-//
-// Background: issue #5 was the inverse of issue #6.
-//
-//   issue #5 (0.2.0–0.2.1): the JVM JAR bundled `org.meshtastic.proto.*`
-//   alongside `org.meshtastic.tak.*`. Consumers that also ran Wire codegen
-//   against the shared `meshtastic/atak.proto` (notably Meshtastic-Android's
-//   `core:proto` module) hit R8 "Type is defined multiple times" errors
-//   during release builds. The fix at the time was to strip ALL proto
-//   classes from the JAR so the consumer's codegen became the single source.
-//
-//   issue #6 (this change): that strategy is ABI-fragile. The SDK's
-//   bytecode REFERENCES the proto classes — `SensorFov.getRange_m()`,
-//   `TAKPacketV2.Builder.payload(...)`, etc. — and those signatures shift
-//   whenever the proto file changes (e.g. when `range_m` flipped from
-//   `uint32` to `optional uint32`, Wire's generated accessor went from
-//   `int getRange_m()` to `Integer getRange_m()`). If the SDK was compiled
-//   against the new proto but the consumer's `core:proto` regenerated
-//   against an older submodule pointer (or vice versa), the runtime
-//   classpath had method signatures the SDK's call sites didn't match,
-//   producing NoSuchMethodError at first invocation.
-//
-// New strategy (#6): the SDK is the single owner of `atak.proto` codegen
-// for ALMOST every type. The JAR ships `org.meshtastic.proto.**` (115
-// classes) alongside `org.meshtastic.tak.**`. Consumers prune the same
-// types from their own Wire codegen and pull them transitively from this
-// JAR. No second codegen, no ABI drift.
-//
-// The two exceptions are `meshtastic.Team` and `meshtastic.MemberRole`.
-// Both enums are referenced from `meshtastic/module_config.proto`'s
-// `ModuleConfig.TAKConfig` (`Team team = 1`, `MemberRole role = 2`), so
-// consumer Wire codegen REFUSES to prune them — pruning would propagate
-// to those fields too, leaving consumer code with no way to read
-// `takConfig.team`. We strip them from this JAR so the consumer's own
-// codegen remains the single source; the SDK's bytecode references to
-// `Team` / `MemberRole` resolve transitively at the consumer's classpath.
-// These two enums are tiny and stable (no field shape to ABI-drift), so
-// the issue #6 ABI argument doesn't apply meaningfully to them.
-//
-// iOS klibs ship all proto classes unconditionally — no R8 step on Apple.
-//
-// See https://github.com/meshtastic/TAKPacket-SDK/issues/6 for the full
-// rationale and the Meshtastic-Android-side prune list.
-tasks.named<Jar>("jvmJar") {
-    // Strip ONLY Team and MemberRole (and their nested ProtoAdapter inner
-    // classes). Everything else in org.meshtastic.proto.** stays in the JAR.
-    exclude("org/meshtastic/proto/Team.class")
-    exclude("org/meshtastic/proto/Team\$*.class")
-    exclude("org/meshtastic/proto/MemberRole.class")
-    exclude("org/meshtastic/proto/MemberRole\$*.class")
-}
-
 // ─────────────────────────────────────────────────────────────────────────────
 // Maven Central publishing via Vanniktech maven-publish plugin.
 // Coordinates are read from gradle.properties: GROUP, POM_ARTIFACT_ID, VERSION_NAME.