toktrace
v0.2.0
Published
Zero-code LLM call tracing and cost tracking
Maintainers
Readme
toktrace
TokTrace is a lightweight Node.js token/cost tracker for LLM calls.
60-second quickstart (no app code changes)
npm install toktrace
npx toktrace install
npx toktrace run -- npm run dev
npx toktrace verifyThat toktrace run command injects --import toktrace/auto so OpenAI/Anthropic/generic HTTP patching is enabled automatically.
Standard quickstart
npm install toktrace
npx toktrace initThen either:
Run your app with auto-instrumentation:
npx toktrace run -- node server.jsOr add one line in app startup:
import { init } from "toktrace"; init();
Dashboard
npx toktrace dashboardOpen http://localhost:4242.
No events yet? Seed sample data:
npx toktrace seedWhat gets captured
- timestamp
- model/provider
- input/output/total tokens
- estimated cost
- latency
- prompt hash (full prompt body is not stored by default)
- environment tag
- tool call metadata (
tool_call_count, serialized tool calls)
Importing from AI coding CLIs
Already using Claude Code? TokTrace can import every LLM call from your local session logs (no proxy, no key, no runtime injection):
toktrace import claude-codeWalks ~/.claude/projects/*/*.jsonl, upserts one event per assistant message with full token breakdown (input / output / cache-read / cache-creation) and cache-aware cost estimation. Re-runs are idempotent — safe to put on a cron. Events are tagged env=claude-code so you can filter or compare to your app-level LLM traffic.
Cost estimates for Claude Code Max subscribers are hypothetical pay-per-token numbers, useful for tracking usage volume and per-project burn but not your actual bill (which is flat).
Useful commands
toktrace init— create local config + database.toktrace install— same as init, plus zero-code run instructions.toktrace run -- <command>— run any command with auto-instrumentation.toktrace verify— validate setup and first-event ingestion status.toktrace seed— insert sample events for instant dashboard preview.toktrace dashboard— launch local dashboard.toktrace import claude-code— import usage from Claude Code session logs.toktrace suggest— generate optimization suggestions.toktrace snapshot create --name "before"andtoktrace snapshot export --name "before".
