Your first question

The first thing to run is the dry run. It boots the whole service exactly as npm run dev would, forces dry run on top of whatever .env says, asks one question, and prints everything that happened. Nothing in it can move money.

npm run dry-run -w @olai/agent -- "Should I trim my BNB position before the weekend?"

With no question after the --, it asks that same default question (packages/agent/scripts/dry-run.ts, DEFAULT_QUESTION).

This page runs the whole path from the command line on purpose, since it is the fastest way to see every step at once with nothing hidden. To watch the same path by clicking instead, open the desk at /app. The welcome card, Your analyst is ready., walks the same three steps: check the rulebook, ask a question, approve or reject what Olai proposes. Type into the box at the bottom or tap one of the three example chips, watch Olai's steps arrive one plain sentence at a time, and answer the proposal card at the end.

What it prints, in order#

1. What it booted with. One line naming the exchange it connected to, the ledger file, and the public address:

Olai booted: dry run, exchange <name>, ledger ./data/olai.db, address http://127.0.0.1:4000

With OLAI_EXCHANGE=auto, the name comes from what you have: the Binance REST key first, then a stored MCP token, then the deterministic stand-in, and the stand-in in a dry run only. buildService refuses to boot a live run without a real connection.

2. The question, back as Olai read it.

3. The thinking, as it happens. The script subscribes to GET /api/events, the same server-sent event stream the desk itself reads (packages/web/src/lib/sse.ts), so this output proves the feed works rather than reaching inside the service for a shortcut. Thinking text and answer text stream in fragments. Each tool call prints twice, once when it is asked and once when it answers:

  asking search_bazaar
  search_bazaar: Bazaar search "BNB wallet flows" under $0.05: 3 payable results
  asking buy_data
  buy_data: ...
  asking read_ticker
  read_ticker: BNBUSDT at ..., ...% over 24 hours

4. The proposal. One summary line, then the model's confidence between 0 and 1, the action as JSON, and one line per risk it named:

Proposal: <one sentence>
  confidence: 0.62
  action: {"type":"order","symbol":"BNBUSDT","side":"SELL","quoteUsd":15,"orderType":"MARKET"}
  risk: <what could go wrong>

The only two actions that can come back are order and hold (proposalSchema in packages/agent/src/brain/proposal.ts).

5. The rulebook's verdict. Allowed or refused, with the rule ids that fired and their plain-English reasons:

Rulebook: allowed (order.needs_approval) This order is above $0.00, so the owner has to approve it.

6. The approval. The script plays the owner and approves, so you see the whole path. The rulebook is checked a second time here against a fresh account snapshot before anything is sent. In a dry run the order stops there: a ledger line marked dryRun: true is written and exchange.placeOrder is never called.

7. Every ledger line the session wrote, with its sequence number, kind, actor, cost where there is one, and its summary:

Ledger lines for this session:
  12 question by owner: ...
  13 discovery by agent: ...
  14 payment.preview by agent ($0.0100): ...
  15 proposal by agent: ...
  16 approval by owner: ...
  17 order.sent by agent: ...

8. Whether the chain still holds.

Ledger chain holds across all 17 lines.

The script exits 0 when the chain holds and 1 when it does not, so it is usable in a check.

Then prove it against the live network#

npm run prove -w @olai/agent

Seven steps against the real Binance Agentic Wallet, the live B402 Bazaar, a live merchant's 402 response, the exchange Olai booted with, and the ledger. Each prints PASS, FAIL or SKIP with a reason. Only step 4 can spend money, and only when .env has both OLAI_DRY_RUN=false and OLAI_PROVE_SPEND=yes. Everything else is read only.

PASS 1. wallet CONNECTED, x402 daily limit $..., quota left $...
PASS 2. Bazaar has N wallet payable listings for "wallet balance" under $0.05, cheapest $... at ...
PASS 3. a real 402 previewed to payment ... with N options, top option is READY_TO_SIGN for $... in ... on chain ...
SKIP 4. OLAI_DRY_RUN is true, so Olai will not sign anything
PASS 5. ... priced BNBUSDT at ...
PASS 6. the ledger hash chain holds across all N lines
PASS 7. this run spent $0.0000 across 0 settled payments (0 signatures claimed)

If something goes wrong#