Token Trace Kit — AI skill for Claude Code

TokenTrace

**One trace from tap to token: the user's tap, the on-device model, your proxy and the Claude API call, on one clock, with no prompt text in any of it.**

Your AI feature is now four systems. The tap happens in SwiftUI. The first attempt runs on-device. When the model refuses or the device can't run it, the request goes through your backend proxy to a vendor model. On 14 Oct 2026 the Claude Platform launched [OpenTelemetry export in beta](https://platform.claude.com/docs/en/release-notes/overview) for Messages API requests, so the vendor side can now land in the same backend as yours. That only helps if the four halves join into one trace, and on a phone they usually don't. The clock is wrong by seconds. The traces you need (refusals, fallbacks, timeouts) are the ones a 5% head sampler throws away. And the obvious attribute to add while debugging is the prompt.

TokenTrace is a Swift 6 library for the client half of that problem: W3C trace propagation, tail sampling on the device, default-deny redaction, cost attribution per feature, clock-skew correction and a bounded offline export buffer. It has no dependencies and runs on iOS, macOS and Linux.

**Demo app:** [token-trace-demo-app](https://github.com/rajatslakhina/token-trace-demo-app). Tap once and see the tap, the on-device refusal, the proxy and the Claude call as one 5-span waterfall on the server's clock. Then flip *Show without skew correction* and watch the proxy span jump in front of the request that caused it.

---

Why this matters

Distributed tracing on the server is a solved problem. On a phone, the four assumptions it rests on are all false:

Server assumption	On a phone	What TokenTrace does
Clocks are NTP-synced to milliseconds	The device clock is off by seconds and can jump mid-request	Durations use the monotonic clock only. A four-timestamp (NTP-style) estimate from your own proxy's responses, anchored to the monotonic clock, moves exported timestamp