Grok Bot Trading Floor: how to run an 8-agent crypto desk in 10 Steps (Full-tutorial)

Cover image

Elon Musk made a repost of the video guide and I decided to write a guide in text format about agents

Every trading bot you have used so far was a script. You wrote a rule, it obeyed the rule, and when the market did something the rule did not cover, it kept obeying anyway.

Grok Bot breaks that. Each bot gets its own computer in the cloud. It opens rugcheck the way you would, logs into gmgn with your session, reads what it finds, and argues with another bot about whether the trade is worth taking.

Which means you stop writing conditions and start running a desk. You stop debugging a strategy and start hiring roles.

This is the 10-step tutorial from installing the app to running eight specialists who pass one trade between each other - and pull you in at the single step that spends money.

01. Install it, then hire a coordinator - not a trader

Grok Bot is a desktop app for macOS and Windows, not a browser tab. You sign in with your Grok or Cursor account and land in something closer to Slack than to a chatbot: named bots down the left, a conversation on the right.

Your first bot should not touch a chart. Hire the coordinator first - the one you message when you do not yet know which specialist owns the question. Everything else reports to it later.

Give it one errand you can check in thirty seconds. Something like: open gmgn.ai, list the ten Solana tokens with the highest hourly volume right now, give me the contract address and a link for each. Then open two of those links yourself and confirm the addresses match.

If it fumbles that, it has no business near a wallet. Fix the basics before you scale the org chart.

02. Fund a wallet you are willing to lose

Make a fresh Phantom or Solflare wallet that has never held anything else. This wallet exists for the floor and nothing else, so when something goes wrong the damage has a ceiling you picked in advance.

Then learn the login handoff, because it is the whole trust model of this product. The bot drives its own cloud browser until it hits a wall - a wallet connect prompt, a Telegram QR, a login form. It hands you the screen. You approve, click done, and it resumes on the same browser instance.

The bot ends up with a session. You keep the secret.

// the rule that does not bend
Never type a seed phrase or private key into a chat message.
Never paste one into a page a bot navigated you to.
Approve wallet connections yourself, in the handoff, every time.

// if a bot ever asks for a key in a message
That is the end of the experiment, not a step in it.

03. Write charters, not prompts

A prompt is a request. A bot is a role, and roles persist - the thread accumulates memory, the bot gets sharper at one domain because you keep coming back to it.

Brief each one the way you brief someone on day one: what they own, what good looks like, and the part everyone skips, which is what they must never do without asking you.

That last block is what lets you close the laptop. Start with the Head of Desk.

You are my Head of Desk.

// what you own
Take my weekly objective and turn it into daily work for the floor.
Delegate in the Trading Floor group chat. Coordinate every handoff.
Keep the funnel: scanned -> cleared -> confirmed -> executed -> closed.
Send me a morning status and an evening report.

// what good looks like
I hear from you twice a day, plus once per decision that needs me.
Every trade has a numbered thread I can read from lead to exit.
Nothing reaches SNIPER without a RISK clearance and two confirms.

// where you stop
You never trade, never sign, never hold keys.
You never let a lead skip a stage, however obvious it looks.
You never raise a risk limit. I do that, in writing, in this chat.
At 5% daily loss you halt the floor and tell me. No exceptions.

04. Hire seven specialists, not one trader

Split by domain, not by task size. A bot that only ever thinks about contract audits gets good at contract audits. One generalist juggling discovery, safety and execution is mediocre at all of them and leaves you no clean thread to read when a trade goes wrong.

Three of them find things:

// SEARCH - alpha hunter
Own: scan 24/7. New pools on pump.fun and gmgn, early liquidity
adds, fresh dev deploys, the Telegram channels I connected, GitHub.
Good: 20-40 leads a day, each with an on-chain reason attached.
Stop: you never recommend an entry. Leads go to RISK, nobody else.

// WHALE - smart money tracker
Own: watch the wallet list I gave you, win rate 60%+. Flag stealth
accumulation, cluster buys, wallets that historically front a run.
Good: you catch accumulation before the chart shows it.
Stop: you confirm or you stay quiet. You never size a position.

// SHILL - sentiment engine
Own: social velocity on X and Telegram. Mention rate, who is
posting, how early. Flag when a name goes from 0 to loud.
Good: you tell me the difference between organic and paid.
Stop: hype alone is never a confirm. No social signal buys anything.

Two of them stop things:

// RISK - contract auditor
Own: audit every lead. Mint authority, freeze authority, LP lock
and duration, owner permissions, tax logic, honeypot behaviour,
holder concentration. Use rugcheck, solscan, dexscreener.
Good: you reject fast and you say exactly which check failed.
Stop: nothing moves past you without an explicit CLEAR from you.

// RUG - sentinel
Own: after every entry, watch the dev wallet and the LP every
minute. Any LP pull, any suspicious transfer, any mint event.
Good: you fire DUMP within seconds, before the chart moves.
Stop: you raise the alarm and hand to EXIT. You never close alone.

Two of them have hands:

// SNIPER - execution
Own: prepare orders on Jupiter or Photon once a lead is cleared
and confirmed. Size, entry, stop, targets, slippage, priority fee.
Good: the order is staged and ready before I read the message.
Stop: you do not sign anything without my literal APPROVE #id.
     Not a summary, not a paraphrase, not "he seemed fine with it".

// EXIT - position manager
Own: every open position. Trailing stops, scaling out, invalidation.
Good: you take profit on the way up, not on the way back down.
Stop: you never add to a position and you never average down.

Build four of these first. Head, SEARCH, RISK, EXIT is already a complete loop - find something, check it, hold it, get out of it. Add the other four once those four have bored you for a full week.

05. Show them the trade once instead of describing it

You can teach a bot a workflow by doing it once while it watches. It saves the routine and runs the same steps later on its own.

This matters more than it sounds, because the parts of trading that eat your evening are tedious to write down and trivial to demonstrate. "Open the pair on gmgn, check the LP lock on rugcheck, look at the top ten holders on solscan, then stage a swap on Jupiter with 3% slippage" is a paragraph to type and forty seconds to show.

Start the recording, then run one full cycle by hand, slowly. Find the token, pull the contract address, check mint and LP and holders, stage the swap, set the trailing stop.

Then record a second one where you reject the token halfway through, and say out loud which check killed it. If every recording you make ends in a buy, you have taught the floor what buying looks like and nothing about what passing looks like. Most of the job is passing.

06. Put them in one room and give them an objective

Make a group chat, call it Trading Floor, add all eight bots. Everything below happens in that one room, which is also what makes it auditable later.

Then hand the room an objective, not a task list. A task list means you already did the decomposition and the bots are typing out your plan. An objective lets the Head split it, which is the reason you hired more than one.

Objective for this week.

Target: 15-20% on the book, no forcing it on quiet days.
Max risk per trade: 1.8% of deposit.
Max daily loss: 5%, then the floor halts.
Universe: Solana memecoins, plus majors if a clean setup shows up.
Quality bar: high. I would rather see 3 leads than 30.

Run the full cycle:
1. SEARCH - fresh leads
2. RISK - clear or reject each one
3. WHALE + SHILL - confirm or pass
4. SNIPER - stage and wait for my approve
5. RUG + EXIT - on watch from the moment we are filled

Morning status in 20 minutes. Work the protocol. No entries on your own.

One thing about that target number: typing 18% does not produce 18%. What the line does is give RISK a size to enforce and EXIT a loss to halt on. Treat it as a boundary, not a forecast.

07. Force one message format on everyone

Skip this step and your group chat turns into eight bots free-associating about a ticker. Impose a format on day one and the same chat becomes a ledger you can read six weeks later.

Every message carries the lead ID. The ID is the thread.

LEAD #047 | 09:14 UTC
Token: $TICKER
CA: 7xKXtg2CW87d97TXJSDpbD5jBkheTqA83TZRuJosgAsU
Source: Telegram @channel + dev wallet deployed 3h ago
Why: LP added 12 min ago at $48k, volume already $120k
Social: nothing on CT yet
-> RISK

RISK CLEARANCE #047
Mint: renounced | Freeze: none | LP: locked 100%, 6 months
Honeypot: no | Top 10 holders: 31% | Tax: 0/0
Score: 9.2/10
Verdict: CLEAR -> WHALE + SHILL

RISK REJECT #047
Failed check: mint authority live, owner can change tax
Verdict: REJECT. Thread closed.

WHALE #047
3 wallets, win rate 71-84%, accumulated $18k in 40 min
Signal: STEALTH ACCUMULATION -> CONFIRM

SHILL #047
Social velocity +340% in 25 min, no influencer cluster yet
Stage: early -> CONFIRM

The repeated contract address is doing real work there. The same string has to appear in the lead, the clearance and the order, so a hallucinated address gets caught by comparison instead of by your wallet.

Post the routing rule once, in the group, and hold everyone to it:

Handoff protocol. Nobody skips a stage.

SEARCH -> RISK only
RISK CLEAR -> WHALE + SHILL only
both CONFIRM -> Head, who briefs SNIPER
filled -> RUG + EXIT immediately
RUG DUMP -> EXIT, then Head, then me

One CONFIRM is not two. A REJECT is final unless I overrule it myself.

This is where bad trades die. The failure mode of a chatty multi-agent system is enthusiasm - two bots agreeing on a feeling and a third one executing it.

08. Draw the approval line at the signature

Grok Bot's premise is that a bot finishes the job and comes back only when something needs you. So you have to define what needs you, because the default the bot picks will not match yours.

Draw the line at reversibility. Anything the bot can undo, it finishes alone. Anything that signs a transaction stops and waits.

// do these alone, always
scan · audit · score · confirm · size · stage the order
draft the exit plan · move a stop tighter · log · report
   everything reversible. do it, log it, do not ask.

// park these for me, always
sign any transaction, buy or sell
move funds between wallets
approve a token spend
change a risk limit or a daily loss cap
add a new source, channel or wallet to the watchlist

// the exception that overrides the above
RUG fires DUMP -> EXIT closes immediately, tells me after.
Getting out is the one irreversible action I pre-approve.

Your side of it is two words:

APPROVE #047

Give it four to six weeks before you loosen anything, and let your own numbers decide. If you approve 4 of the 9 setups the floor brings you, the floor is wrong more often than right, and automating that only means being wrong faster. When your approve rate sits high for a month, drop the manual gate on one narrow case - small size, one type of setup - and watch it.

Keep a kill switch and test it while nothing is wrong:

On "STOP ALL": close every open position at market, cancel every
staged order, stop the cycle, report what you closed and at what.
No confirmation needed. No questions. Then wait for me.

09. Turn the boring parts into routines

Two routines carry most of the load. Set them up conversationally, no workflow builder.

Position watch, on RUG and EXIT:

Every 15 minutes, check every open position: price against entry,
LP status, dev wallet movement, volume trend, stop distance.
Message the group only if something changed or a stop moved.
Silence means everything is where I left it.

Evening report, on the Head, scheduled daily:

=== EVENING REPORT [date] ===
Scanned by SEARCH: XXX
Cleared by RISK: XX      (rejection reasons, top 3)
Confirmed by WHALE+SHILL: XX
Executed by SNIPER: X
Open now: X

PnL today: +$XXX / -$XXX, after fees and priority fees
Win rate: XX%
Best: $TICKER +XXX%   Worst: $TICKER -XX%
Risk: within limits / breached at HH:MM

One thing that worked. One thing to change tomorrow.

Read the funnel before you read the PnL. 300 scans producing 2 clears means SEARCH is shovelling noise at RISK. 40 clears producing 1 confirm means your confirmers and your hunter disagree about what a good setup is, and one of the three charters needs rewriting.

Have the Head keep a file on its cloud computer too:

Create trading_log.md and append one row after every close:
date | ticker | CA | entry | size | exit | PnL | reason for exit
Never edit an old row. Add a correction row instead.

10. Review weekly, prune ruthlessly

Automation rots quietly. gmgn ships a layout change, a selector breaks, and SEARCH starts reporting an empty scan every hour while sounding perfectly calm about it. Because the floor runs while you sleep, that failure can go three weeks before you notice.

Book fifteen minutes on a Sunday and ask the floor to report on itself:

For the past 7 days, per bot:
- how many messages you sent, and how many were signal
- what you flagged that turned out wrong, and why
- anything you could not check because a page or login failed
- anything you handed me that I never answered

Then tell me which agent earned its seat least this week, and why.

Then verify by hand. Pull two contract addresses out of the log and confirm them on solscan yourself, character by character. An address that is one character off is the single most expensive hallucination available to this stack, and no bot in the group will catch it for you.

Now the part that gets left out of the screenshots. The "+$13k this week, 7 autonomous trades" posts come from good weeks in a hot memecoin season, published by people who do not publish the other weeks. What this floor does reliably is research, filtering, monitoring and preparation - the work that otherwise has you holding fourteen tabs open at 2am. The version where you sleep and the money arrives is a much later stage of maturity, and it still needs a market willing to cooperate.

Conclusion

You stop being the one refreshing the chart at 2am.

Every AI trading tool before this put a model in a window and left the clicking to you. Grok Bot moves the work. The bots hold the computer, the sessions, the memory and the patience. You hold the wallet and the decisions.

Smaller change in technology than it sounds, much bigger change in habit. The skill stops being what do I buy and becomes what am I willing to let run unattended, and where does its authority end.

CA: GXdbB581sDfMuhwJsQUksbvhj9yPTosaxhaKuHiTpump

That is a risk management question, and you were always going to have to answer it.