All tests run on an 8-year-old MacBook Air.
Most AI integration tutorials assume you're paying for API access.
HiyokoLogcat is built entirely on Gemini's free tier — and designed so users bring their own free API key. Here's what's possible, what the limits are, and how to design around them.
The free tier numbers (as of 2026)
Gemini 2.5 Flash Preview:
- 15 requests per minute (RPM)
- 1,000,000 tokens per day
- 250 requests per day
For a developer tool used intermittently — click diagnose, read, fix, move on — these limits are generous. A typical diagnosis uses 3,000–6,000 tokens. At that rate, you'd need to run 166+ diagnoses to hit the daily token limit.
Design for the free tier from the start
Keep requests small. Don't send 500 lines when 100 will do. Every token saved is headroom for more requests.
Don't auto-trigger. Never send to the API automatically — only on explicit user action. Auto-triggering on every error would burn through the RPM limit instantly.
Cache results. If the user closes the diagnosis overlay and reopens it, serve the cached result. Don't make a new API call for the same log line.
use std::collections::HashMap;
pub struct DiagnosisCache {
cache: HashMap, // log hash → diagnosis
}
impl DiagnosisCache {
pub fn get(&self, log_hash: &str) -> Option<&String> {
self.cache.get(log_hash)
}
pub fn insert(&mut self, log_hash: String, diagnosis: String) {
// Keep cache bounded
if self.cache.len() > 50 {
self.cache.clear();
}
self.cache.insert(log_hash, diagnosis);
}
}
50-entry cache. Old entries cleared when full. Simple and effective.
The "bring your own key" model
Asking users to get their own free API key has an unexpected benefit: they feel ownership over the AI feature.
"I set up my own Gemini key" is a different mental model than "the app has AI built in." Users who set up their own key understand the rate limits, understand that it's a free service, and are more forgiving when it's occasionally slow.
The setup friction (2 minutes to get a key from Google AI Studio) filters out users who aren't genuinely interested in the feature.
What free tier can't do
Batch processing. If a user wants to run AI analysis on 100 error lines at once, the free tier can't handle it without significant delay. Design your UI to discourage bulk AI use.
Real-time analysis. Don't try to send every log line to the API as it streams in. That's a 15 RPM limit burning in seconds.
High-volume production use. If your tool gets popular and thousands of users are hitting Gemini simultaneously, they share the free tier limits per API key. This is actually fine — each user has their own key.
The honest verdict
For a developer tool with intermittent AI use: the free tier is completely sufficient. I've never hit the daily limit in normal use.
Build for the free tier from day one. It keeps your costs at zero, your users' costs at zero, and forces you to design efficient AI interactions rather than spamming the API.
HiyokoLogcat is free and open source → github.com/hiyoyok/HiyokoLogcat
X → @hiyoyok
Top comments (0)