Understand what your AI is actually doing.

A lightweight, open-source collector for AI agents.

Every model call, every agent run, every Claude Code session is a complete record of what your AI did and why. HypAware captures that record and lands it in a bucket you own as Apache Iceberg, so you can finally answer where the spend goes and what is actually working. It runs on your own machine, it is free, and the data never passes through a vendor you don’t control.

$ npx hypaware   # interactive setup, then capturing in under a minute

Get set up↓ Source on GitHub↗

1 Why HypAware

You put a few hundred developers on Claude Code and Codex, billing moved to usage, and the bill keeps climbing. None of that would bother you if you knew it was working, but you can’t answer the basic questions: where the tokens go, which prompts and tools succeed, and which ones quietly loop and burn context. Anthropic hands you a bill, not the session data. The surveys show what people remember. The answer to every one of those questions is in the logs, and almost no one is collecting them.

HypAware is the first step. It records the real traffic so the questions become queries:

Where is spend going? Token cost broken down by developer, repo, and workflow, from real sessions rather than a single line on a bill.
Which prompts and tools work? Which ones succeed, and which quietly loop and burn context, so you can cut what isn’t working.
How is the team using AI? Patterns across every agent and developer at once, not a survey and not one machine.

2 What you get

Your data, your bucket

Logs land in object storage you own as open Iceberg files. No proprietary database in the path, nothing held hostage by a vendor, and you can delete or move it whenever you want.

Open source, free to run

HypAware is MIT licensed and free. A single developer can run the whole thing locally, and the bundled plugins cover the common clients with no external install.

Captures what you already use

Claude Code, Codex, raw Anthropic and OpenAI API traffic, and OpenTelemetry logs, traces, and metrics. One queryable row per content block, normalized to Iceberg.

No heavy backend

It runs as a small daemon on your own machine. Recordings and cache stay local until you choose to ship them, with no always-on server to keep warm between queries.

Query with what you have

Run hyp query sql against captured sessions, or point any engine at the Iceberg tables. The files are standard columnar Parquet, readable by anything.

Scales from laptop to fleet

Start on one machine, then push HypAware to a fleet of developer laptops via MDM so every session lands in one company bucket. Same collector, same format.

3 Get set up

On a terminal, one command starts an interactive walkthrough that picks your sources, writes a minimal config, installs the daemon, attaches your clients, and starts capturing. It finishes by printing a first query to run.

Run the installer. On a TTY this launches the walkthrough.
```
$ npx hypaware
```
Pick your sources and retention. Choose any of Claude Code, Codex, raw Anthropic, raw OpenAI, or OTEL, then a retention window (30 days by default). HypAware installs the persistent daemon and attaches the clients you selected.
Confirm it’s capturing. One command shows config path, daemon state, attach results, and a recent-error count.
```
$ hyp status
```

Query what you captured. Captured traffic is queryable straight away.

$ hyp query sql "select count(*) from ai_gateway_messages"
$ hyp query sql "select content_text from ai_gateway_messages where role = 'user' limit 10"

For unattended installs (CI, scripted bootstraps, dotfiles) skip the picker with explicit flags:

$ hyp init --yes \
    --source claude --source otel \
    --client claude \
    --export local-parquet \
    --retention-days 30

Full documentation↗ Source on GitHub↗

4 Part of HypStack

HypAware is the collection step. Once your logs are landing in a bucket as Iceberg, the rest of HypStack reads them: SQL, full-text, and vector search run client-side straight from object storage, with no cluster to scale and nothing idle between queries. And in Hyperparam, a model reads the same logs and hands back where tokens are wasted and the exact prompt and skill changes that would fix it. Collection is open and free; the value compounds as the data accumulates.

HypStack↗ Hyperparam↗