unt and paying $99/year for the privilege.

It’s a desktop app for macOS (Apple Silicon only for now). You import Chase CSV files, everything goes into local SQLite, and transactions get categorized automatically. There are 50+ built-in rules, learned rules that improve as you correct them, and an optional local LLM via llama.cpp. You bring your own GGUF model. No API calls, no cloud, no telemetry.

Stack: Tauri 2.0 (Rust + React), SQLite, llama.cpp bindings. One thing I haven’t seen in other tools: every transaction shows why it was categorized (rule name, confidence score, or “manual”). You can set budgets, filter/search, and export to CSV.

Free and MIT licensed. Limited right now: only Chase CSV, only macOS. I wanted to ship and see if anyone else cares about this approach.

GitHub: https://github.com/Humanji7/ledgr Download (macOS ARM): https://github.com/Humanji7/ledgr/releases/latest18:T7a4,Hey HN, I built Cobalt, an open-source testing framework for AI agents and LLM apps.

Most eval tools (Braintrust, Arize, LangSmith) want you to live in their UI. Dashboards, manual reviews, clicking through results. That's fine for exploration, but it doesn't catch regressions. We needed something that runs in CI like any other test suite, lives in code, and fails the build when quality drops.

  npm install @basalt-ai/cobalt
  npx cobalt init
  npx cobalt run

Write experiments as code:

  import { experiment, Dataset, Evaluator } from '@basalt-ai/cobalt'

  const dataset = Dataset.fromLangfuse('support-tickets')

  experiment('support-agent', dataset, async ({ item }) => {
    const result = await myAgent(item.input)
    return { output: result }
  }, {
    ev

Why Boys Are Behind in Reading at Every Age

Comments