How the desk works

What this is

A command center for Robinhood Chain. It reads the chain and the market, forms opinions, and lets you rehearse trades on a paper book. It is a research desk: it can quote a swap down to the route and price impact, but it never signs one.

It is not a wallet, not a custodian, and not an exchange. Nothing you do here moves real funds.

The one box

Everything happens in the chat at /os. Ask in plain English; the desk picks the tools, calls them against live sources, and writes a brief. A single question can fan out to several tools. Asking for a risk scan typically pulls the token dossier, liquidity, holder quality and recent flow before it answers.

You will see each tool appear as it runs, with its status, and you can expand any of them to read the raw result the model was given. Nothing is hidden behind the answer.

Two engines

The desk runs on two models, and the split is strict. It is not a routing preference, it is a boundary.

The desk model is ours

0cto-1.2 is not a hosted API. It is a small model trained here, on this desk’s own tool surface — the same 37 schemas the kernel sends, in the same shape it sends them — so it was trained in the configuration it actually runs in, rather than taught to be generally clever and handed a toolbox afterwards.

On the desk’s own evaluation — 310 cases covering routing, write actions, refusals, multi-turn context, ambiguity, stale windows and failed data sources, graded against this desk’s rules rather than another model’s opinion — it scores 96.8% at temperature 0, averaged over repeated runs rather than quoted from the best one. On the narrower suite the previous hosted model was measured against, it answers 97.9% correctly to that model’s 89.3%. It weighs 2.7 GB and runs on a single GPU in the same box as the node.

Which is the point: the desk answers without renting anyone’s model by the token, and what you type stays on the machine.

Replies are always written in English, whatever language you ask in and whatever the tools return.

What to ask

These are examples, not commands. There is no syntax to learn. Anything in this shape works.

• The tape“Give me the desk this morning.”
• Quotes“Quote 1 ETH to USDG on Uniswap v4.”
• Radar“What pools just launched?”
• Risk“Scan this contract for rug and holder quality.”
• Compare“NVDA vs SPY vs HOOD.”
• Wallets“What is this wallet holding?”
• Decode“Explain this transaction hash.”
• Yield“Which v4 pools are earning real fees?”
• Positions“Are my LP ranges still earning?”
• Social“What narrative is soaking up attention?”
• Paper“Buy $500 of NVDA on paper.”
• The book“How is my paper book doing?”
• Watchlist“Add SPY to my watchlist.”
• Rules“Alert me if NVDA drops 10%.”
• DCA“Simulate $50 into ETH weekly.”
• Rebalance“Rebalance me 60/30/10.”
• The cage“Hunt this GitHub repo for bugs.”

Every tool

The desk has 37 tools: 29 that read live market and chain data, and 8 that read or change your own saved state. This list is generated from the running code, so it is never out of date.

Your saved state

Three things persist between sessions: your paper book, your watchlist, and your automation rules. The desk can read and change all three by conversation.

All of it is stored in your own browser under the localStorage key octo-os. There is no account and no server-side copy. Clearing site data resets the desk; a different browser is a different desk.

The paper book

Starts at $10,000 in simulated cash. Fills are priced off live Uniswap v4 quotes at the moment you ask, so the book tracks real prices, but no order is ever routed. A buy that exceeds your cash, or a sell larger than the position, is refused rather than clamped silently.

Automation rules

Rules are reminders, not executors. Adding one records the condition you care about. It cannot place an order, sign anything, or move funds, because the desk has no key with which to do so.

Safety model

• Watch-only trading. No seed phrase, no private key, and the desk never signs a trade, swap or transfer on your behalf. Connecting signs a text message, not a transaction. The single exception is paying for a cage hunt, which you initiate and approve yourself. See The cage. Any address you research is simply read.
• Paper first. Routes and quotes are live and real. Fills land on the paper book. Nothing reaches a validator.
• No invented numbers. Prices, slots, holder counts and risk verdicts come from tools. If a tool fails, the desk says so rather than guessing.
• Memecoins are entertainment-grade risk. The desk is instructed to say “avoid” plainly when a token looks hostile, not to hedge.

Access & rate

The landing page and these docs are open. The desk itself needs a sign-in, by wallet or by email. With a wallet you sign a one-time message, and the desk reads your $0CTO balance to set how fast you can ask.

The signature is over a plain text nonce and is single-use. A captured signature cannot be replayed. What it proves is that you control the wallet, nothing more.

The rate is also per connection. Every account signed in from one address draws from a single allowance at the tier it presents, so creating more wallets or email accounts does not add prompts.

Balances are re-read periodically rather than on every prompt, so topping up takes up to a minute to raise your rate.

Where data comes from

Chain reads go to a dedicated node where one is configured, with redundant fallbacks behind it. Whether the desk is on a private node or a public RPC is shown on the command panel; the endpoints themselves are not published.

Routes

Earlier builds shipped a page per desk: radar, paper, yield and the rest. Those routes still resolve, but they are no longer part of the product surface; the chat covers all of it.