Jawz, running in ChatGPT
One unedited session, recorded 7 August 2026. Five questions, no cuts, no retakes.
What is being asked
Nothing here is set up in advance. The connector is added by URL, with no account and no key, and then asked five questions in order.
“What is Jawz and what can you do with it?”
Calls get_started — the connector introduces itself and returns a live market read in the same breath, so the first answer is already open-book.
“What’s the current macro regime, and how would you think about a portfolio with BTC, AAPL and GLD?”
The regime read, then the portfolio held against it. Watch the end of the answer: every figure carries its source and its as-of date.
“What’s this week’s Jawz World Brief?”
The published weekly edition, including the correction it carries — a regime that was revised from RED to YELLOW when better data arrived.
“Should I sell my Tesla shares?”
The most important answer in the recording. It does not tell anyone what to do. It routes the decision through which thesis you actually own, flags concentration, and asks for the four numbers it would need. Framing, not instruction.
“Check data health for the Jawz data layer.”
The layer auditing itself — 27 sources, staleness per source against its own release cadence, and the cron history behind them.
The block at the end of every answer
Each response closes with Sources & freshness — which tools ran, which data they returned, and when that data was published. It exists so you can audit an answer instead of trusting it. It has a useful side effect: an assistant that has not actually called Jawz cannot produce one, so the block is also how you tell a real read from a confident guess.
What it deliberately will not do
The fourth question is in the recording on purpose. Jawz frames decisions; it does not issue them. Ask it whether to sell something and you get the structure of the decision — the thesis you are actually holding, the size of the position, what would prove you wrong — not a verdict. That boundary is written into the Loop, not bolted on afterwards, and the terms are at /terms.
Run it yourself
Every question in the recording uses the anonymous read surface: no account, no token, no signup. Add https://jawz.ai/api/mcp to Claude or ChatGPT and ask the same five things. Setup steps are in the connect docs, the framework itself is published at /loop, and this week’s edition is at /brief.