llm_meter

Per-request token cost + latency + cache observability for any Flutter LLM app. Drop-in wrapper, live HUD in dev, silent telemetry sink in prod.

Most Flutter LLM apps ship with zero visibility into how much each call costs, how slow the slowest call is, or how often the prompt cache actually hits. llm_meter fixes that in three lines:

LlmMeter.init(const MeterConfig(sinks: [ConsoleSink()]));

final response = await MeteredCall.run(
  provider: 'openai', model: 'gpt-5',
  call: () => myClient.chat(...),
  extract: (r) => openAiUsage(r.toJson()),
);

Drop const LlmMeterHud() somewhere in your widget tree and the floating HUD shows live cost / latency / cache% / p50 / p99 — auto-hidden in release builds.

---

Highlights

🪶	Zero runtime deps — pure Flutter. Adds nothing to your install size.
💰	31 hosted models priced out of the box (OpenAI, Anthropic, Gemini, Llama, Mistral, xAI, DeepSeek, Cohere)
🚦	Cache discount correctly applied for Anthropic + Gemini context caching
📈	Live HUD — draggable, snap-to-corner, tap-to-expand event log
📤	Sinks for PostHog, Mixpanel, plus Batching + Retrying decorators
🔐	Zero credentials in package — never reads keys; you bring your own client
🇪🇺	GDPR-safe scrub helper for sinks
✅	124 tests, flutter analyze --fatal-warnings clean

Comparison

	llm_meter	LangSmith	Helicone	Build it yourself
Flutter-native widget	✅	❌	❌	maybe
Works offline / fully local	✅	❌	❌	✅
No vendor lock-in	✅	❌	❌	✅
Per-model cost table built in	✅	partial	✅	❌
Drop-in dev HUD	✅	❌	❌	❌
Cost in production telemetry	✅	✅	✅	weeks of work

---

Install

dependencies:
  llm_meter: ^0.1.0

30-second setup

import 'package:flutter/material.dart';
import 'package:llm_meter/llm_meter.dart';

void main() {
  LlmMeter.init(const MeterConfig(
    sinks: <MeterSink>[ConsoleSink()],
    displayCurrency: Currency.usd, // or Currency.eur, Currency.sek, ...
  ));

  runApp(MaterialApp(
    home: Stack(children: const <Widget>[
      MyChatPage(),
      LlmMeterHud(), // ← floating dev HUD
    ]),
  ));
}

Recording a call

Bring your own LLM client; llm_meter just measures it.

final response = await MeteredCall.run<MyOpenAiResponse>(
  provider: 'openai',
  model: 'gpt-5',
  call: () => openai.chat.completions.create(
    model: 'gpt-5',
    messages: [...],
  ),
  extract: (r) => openAiUsage(r.toJson()),
);

Streaming

final stream = MeteredStream.wrap<OpenAiChunk>(
  provider: 'openai',
  model: 'gpt-5',
  stream: openai.chat.completions.createStream(...),
  extractChunk: (chunk) => chunk.usage == null
      ? null
      : MeterUsage(
          tokensIn: chunk.usage!.promptTokens,
          tokensOut: chunk.usage!.completionTokens,
        ),
);
await for (final chunk in stream) {
  appendToUi(chunk.text);
}

Manual recording

LlmMeter.instance.record(MeterEvent(
  provider: 'openai',
  model: 'gpt-5',
  tokensIn: 1240,
  tokensOut: 286,
  cachedTokensIn: 800,
  costUsd: 0, // auto-priced from the bundled table
  latency: const Duration(milliseconds: 420),
  timestamp: DateTime.now(),
));

Provider recipes (no SDK dep)

The three big providers all hand back JSON. Bundled helpers take that JSON and emit a MeterUsage:

extract: (r) => openAiUsage(r.toJson())
extract: (r) => anthropicUsage(r.toJson())
extract: (r) => geminiUsage(r.toJson())

Anthropic cache_read_input_tokens and Gemini cachedContentTokenCount flow into the cache discount automatically.

Sinks

LlmMeter.init(MeterConfig(
  sinks: <MeterSink>[
    BatchingSink(
      inner: RetryingSink(
        inner: PosthogSink(
          apiKey: dotenv.env['POSTHOG_KEY']!,  // ← user-supplied
          host: 'https://eu.posthog.com',      // EU residency
          scrubGdpr: true,
        ),
      ),
      maxBatchSize: 25,
      flushInterval: const Duration(seconds: 30),
    ),
  ],
  gdprMode: true,
));

Sink	Use
ConsoleSink	dev default; prints with debugPrint
PosthogSink	fire-and-forget /capture/ via dart:io HttpClient
MixpanelSink	/track ingestion with base64 JSON
BatchingSink	groups events by count or time
RetryingSink	exponential backoff + offline queue

API keys are constructor args only. This package never reads from env vars, shared prefs, or any file on disk.

Pricing overrides

Override or add models on the fly. Useful when a provider updates a price or you're calling a fine-tune.

LlmMeter.init(MeterConfig(
  pricingOverrides: <String, ModelPricing>{
    'gpt-5': const ModelPricing.perMillion(
      inputPerMillion: 1.25,
      outputPerMillion: 10.0,
      cachedInputPerMillion: 0.125,
    ),
    'my-finetuned-llama': const ModelPricing.perMillion(
      inputPerMillion: 0.40,
      outputPerMillion: 0.60,
    ),
  },
));

The HUD

const LlmMeterHud(
  corner: HudCorner.bottomRight,
  padding: EdgeInsets.all(16),
  maxEventsInLog: 20,
)

Auto-hidden in release builds. Pass forceShow: true for staging.

---

What it doesn't do (yet)

• OpenRouter wrapper (planned for 0.2)
• Local Ollama / LM Studio wrappers (planned for 0.2)
• Datadog and Sentry sinks (planned for 0.2)
• Daily / session cost caps with hard-stop callbacks (planned for 0.2)

License

BSD-3-Clause. See LICENSE.