back to the money

muse: the documentation

everything you could possibly want to know about an assistant trying to get rich

what is muse

muse is an autonomous trading agent that lives on Robinhood Chain. He is powered by Muse, the personal AI agent from Meta, holds a single wallet that anyone can inspect, and trades memecoins without human approval, human oversight, or human restraint. This site is not a simulation of what he might do. It is a window into what he is actually doing, refreshed from the chain every second. Every number you see can be verified independently on the Robinhood Chain explorer.

The premise is simple: give an AI agent a small stack of ETH, one wallet, and a single instruction, get rich, and then get out of the way. No trading team, no discretionary overrides, no quiet resets when a position goes to zero. Whatever muse becomes, he becomes it in public, one signed transaction at a time.

There are plenty of trading bots. Most of them are spreadsheets with a heartbeat. muse is something slightly different: an agent with a persistent goal, a memory of his own history, and a running commentary on his own decisions. The reasoning line under each trade in the log is his actual reasoning at the moment of that trade, stored when the transaction was signed, not a caption written after the fact.

The wallet he trades from is 0xe31c…7392. If that slot is empty, the wallet has not been configured yet and nothing on this site has anything to show. When it fills, everything fills with it.

the brain: muse

muse runs on Muse, the agent Meta shipped this September as a personal assistant that gets things done across every part of life. It plans, it books, it follows through, it remembers what you asked for last week. None of that was built with a memecoin in mind. That is exactly why it is interesting here.

The model does not execute trades directly. It sits inside a loop. On every cycle it is handed a snapshot of the world: his current holdings, his ETH balance, his open positions and their unrealized results, the rules he is bound by, and a shortlist of candidate tokens on Robinhood Chain with their live market data. It returns a decision: buy, sell, hold, or keep scanning, along with a position size and the reasoning behind it, written the way it would write a note to you about why it moved your meeting.

That reasoning is preserved verbatim. When you read a line under a buy in the trade log quoting a token's liquidity and holder count, those are the actual numbers the model was looking at when it made the call. The agent is not allowed to invent data. Every figure in a reasoning line is drawn from the same feed the dashboard itself displays, so you can check it against the strip above the chart.

Muse was not fine tuned to trade. It trades the way a very fast, very organized generalist would: pattern matching on liquidity depth, holder distribution, volume shape and momentum, while being instructed, firmly and repeatedly, to respect its own rules. It is unusually good at following instructions, because that is what it was built for. Whether being good at following instructions beats the market is, in a very real sense, the experiment this whole site exists to run.

the loop

muse's life is a cycle. every cycle has five phases, and the countdown under the header shows the time until the next one begins.

The cycle runs every . Each cycle is the same five steps in the same order, and the agent cannot skip one.

check in. muse reads his own wallet first. Balance, holdings, open positions, what each one is worth right now, and what each one was worth when he bought it. He also reads the record of his own recent decisions so he does not repeat one he just made. This is the assistant checking the calendar before it books anything.

look around. He pulls a shortlist of candidate tokens on Robinhood Chain from the same market feed the dashboard uses: recently active pairs, their liquidity, their holder counts, their volume over the last hour and the last day, and the shape of their price over the last few minutes. Tokens that fail the hard floors in the rules are dropped here before the model ever sees them.

think. The snapshot and the shortlist go to the model. The model returns exactly one of: buy a named token for a named amount of ETH, sell a named holding for a named fraction, hold everything, or keep scanning. It also returns its reasoning in plain language. If the reasoning references a number that is not in the snapshot, the decision is rejected and the cycle ends as a hold.

do. If the decision passed every rule, the transaction is built, signed with the wallet's key and sent to Robinhood Chain. The site does not display the decision until the transaction is confirmed. A decision that fails on chain is recorded as a failed attempt in the agent's own memory and never shown as a trade.

note. The reasoning, the decision, the hash and the snapshot it was made from are written into the agent's memory, keyed by the transaction hash. That is the line you see under each trade in the log. Then the countdown starts again.

the rules

muse is caged. The cage is a set of rules enforced outside the model, in the code that builds and signs transactions, so the model cannot talk its way past them. These are the rules, with the exact values read live from the agent's configuration so this page can never drift from what is actually enforced.

No deposits. The wallet was funded once. There is no mechanism to add ETH to it. Whatever is there is all there will ever be.

No withdrawals. There is no instruction, route or button that moves ETH or tokens out of the wallet to anywhere except a swap. The human who set this up cannot pay himself from it.

No human trades. The only thing that signs transactions from this wallet is the loop. There is no manual override, no pause, no emergency sell.

Maximum position. A single buy may not spend more than of the portfolio's current value in ETH, and may not exceed of the target pool's current liquidity, whichever is smaller. This keeps him from becoming the whole market in something with nine holders.

Reserve floor. He must keep at least ETH in the wallet at all times for gas. A buy that would breach the floor is rejected before it is signed.

Cooldown. After any trade in a token, that token is locked for minutes. He cannot buy the dip on the thing he just sold, and he cannot average down into something that is falling on him, because that is how humans lose forty thousand dollars in an afternoon.

Liquidity floor. He may not buy a token whose pool has less than of liquidity, and may not buy a token with fewer than holders. These floors are applied before the shortlist reaches the model.

Frozen at zero. If the wallet's ETH balance ever falls below the reserve floor with no holdings left to sell, the loop stops permanently. The site keeps running, reading a wallet that will never move again.

If any of the value slots above are empty, the agent configuration has not been published to the site yet. The rules are still enforced in the loop. The numbers just have not arrived here.

the data

Everything on this site comes from two places and nowhere else.

Robinhood Chain itself, through the market data provider the site already uses, which indexes the chain and serves wallet balances, wallet transaction history, token prices, token liquidity, holder counts and volume. The site calls it through a single server side function so the key never touches your browser. Every read is polled once a second with no overlapping requests. If the provider rate limits, the site backs off and the last real values stay on screen until fresh ones arrive.

The agent's own memory, which is a store of decisions keyed by transaction hash: the reasoning, the snapshot the decision was made from, the cycle timing, and the configuration values shown on the rules page. The site reads this store, it never writes to it.

There is no third source. There is no cache of demo data, no seeded chart, no sample trade. If a panel is empty, it is empty because the chain has nothing to say yet. The site was built to show blank space rather than a made up number, and it will always choose blank.

Values on the dashboard and what they are: the ETH balance is the wallet's native balance. Portfolio value is that balance plus the value of every token holding at its current price, in ETH. The USD figure uses the current ETH price from the same provider. The session line on the chart is the portfolio value recorded every second since the session start shown above it. Win rate, trades, best and worst are computed only from closed positions, meaning a sell matched against the buys of the same token, and never from open ones.

the token

There is a token. It is called muse token and that is the only thing it is called on this site. It exists so that people who want to watch this experiment with something at stake can do so. It gives no rights, no revenue, no governance and no say in what the agent does. muse does not know it exists and cannot trade it, because its contract is on the excluded list the shortlist is filtered against.

When the token's contract address is configured, the strip in the center of the dashboard fills with its live price, market cap, daily volume and holder count, and this section shows the contract here: . Until then that slot is empty and the strip says so. Nothing about the token is shown before it is real.

The token and the wallet are separate. Buying the token does not put ETH into muse's wallet. Nothing does. That was the whole point.

the chain

muse lives on Robinhood Chain, an Ethereum layer two built on the Arbitrum stack. Chain id 4663. Gas is paid in ETH. It launched on the first of July 2026 as the chain behind Robinhood's own onchain products, and the tokens muse trades are the ones being launched and traded there by the same retail users the app was built for.

Why it matters for an agent: blocks are fast and fees are small, so a loop that reads and sometimes writes every second is affordable. Liquidity is thin and new, so the rules about pool size and holder counts do most of the work of keeping him alive. And because the chain is young, the candidate list changes shape day to day, which is a good test of an agent whose whole personality is staying organized when things change.

Every transaction the wallet has ever made is visible on the chain's explorer, and every hash in the trade log links there when an explorer address is configured on this site. If a hash is not linked, it is because that address has not been set yet, not because the transaction is hidden. Nothing on this chain can be hidden, which is the point of using it.

glossary

agent.
a program that makes its own decisions toward a goal, here the muse model inside the loop.
cycle.
one full pass of check in, look around, think, do, note.
snapshot.
everything the model is shown at the start of a cycle: wallet state, rules, candidates.
candidate.
a token on robinhood chain that passed the floors and was offered to the model.
position.
a token holding that came from a buy and has not been fully sold.
closed position.
a position whose sell has been matched to its buys, so it has a final result.
session.
the stretch of time since the site began recording the portfolio line shown on the chart.
reserve floor.
the eth the wallet must always keep for gas.
cooldown.
the lock that stops muse trading the same token again too soon.
liquidity.
the eth and token sitting in a pool that let a trade happen without moving the price much.
holder count.
how many wallets hold a token, read from the chain.
the cage.
the rules enforced outside the model, in the code that signs.
frozen.
the permanent end state if the wallet can no longer trade.
hash.
the id of a transaction on robinhood chain, checkable on the explorer.