AGANAL - See what your AI coding agents actually did

Hey Product Hunt 👋

I build my own product, and these days most of the work runs through AI coding
agents — Claude Code, Codex, Cursor — plus a pile of custom MCP servers I've
wired up. Which is great, until a session goes sideways and you're left
wondering: which tool ate all the context? Why did this run balloon past 600K
tokens? Is that MCP actually pulling its weight, or just poisoning every prompt?

With a lot of custom MCPs it's genuinely not trivial to see what's happening, and
reading raw JSONL logs to find out is miserable. So I built AGANAL — a native
macOS app that reads the session logs your agents already leave on disk and turns
them into analytics. One shared model across every provider, so the same
dashboard works whether the run came from Claude Code, Codex, or opencode.

Where it's earned its keep: finding the MCPs and MCP functions that quietly
poison a session — the tools that fire on every turn, dump huge results into
context, or never get used at all. Once you can see tool usage, token cost, and
context-by-category per run, it's obvious what to cut. Trim a noisy MCP (or a
couple of its functions) and the agent gets faster, cheaper, and sharper.

What's inside:
- 📊 Tokens over time, tool-call breakdowns, context-window pressure, retries
- 🔎 A filterable event stream — with the raw JSONL underneath
- 🤖 A CLI and an "Analyse with Agent" tab that hands any session to an LLM
to summarize / find errors / review cost
- 🔒 100% local — it only reads files already on your disk. No account, no
upload, no telemetry.

It's free and open source (Swift/SwiftUI). Fun bit: the app's own
agent-analysis feature caught a real token-counting bug *in the app* while I was
testing it 🙂

Would love your feedback — especially which signals you'd want for spotting
context bloat, and which agents to add next.

→ Free download + source: https://aganal.the-ihor.com

Finally a way to see what claude code and cursor are actually doing under the hood. Token spikes and context pressure graphs were eye opening on my own sessions. Runs smooth on mac.

Love that it just reads the logs already on disk instead of hooking into the tools themselves. That kind of passive design is exactly right for something local and private, and shipping it open source on top is a nice flex.

Finally something that surfaces context pressure across my Claude and Codex sessions without sending data anywhere, that privacy angle sold me. The token and retry breakdowns are exactly the metric I didn't realize I needed.