fly-trader

fly-trader — distribution

A fruit-fly connectome that trades Solana memecoins with your money.

⚠️ Read this first

This software signs real transactions from a wallet you fund. It buys tokens that graduated from pump.fun, holds them for about two hours, and sells. It can lose everything in that wallet, quickly. The model was trained on a few months of data; the market it trades is adversarial and changes. Nothing here is a claim of profit, and nothing here is financial advice. Run it only with money you can lose entirely, on a wallet used for nothing else. MIT licence — no warranty of any kind. See LICENSE.

This is the ready-to-run branch: the trading engine, its console, a trained fly, the selector that taught it, and the connectome they run on. The training pipeline lives on master.

What it does

• Every minute it scores every PumpSwap token that graduated from pump.fun at least six hours ago, from live 1-minute candles (a token needs that much history before its features mean anything).
• The fly — the FlyWire Drosophila central brain (41,756 neurons) as a rate network on its real signed synapses — predicts each token's net return over the next 120 minutes. It was taught by a gradient-boosted selector; it then keeps learning from what its own trades earn, through plasticity at its mushroom body's KC→MBON synapses.
• It buys when the predicted return clears its own buy line, sizes the position from measured certainty (a quarter of the growth-optimal bet, capped), holds 120 minutes, sells.
• It starts on paper. After 20 closed paper trades it switches itself to the wallet and stays there. Before that, and whenever the wallet is unfunded or a key is missing, it trades paper and says so on the console.
• Rails: 20 % below the wallet's peak it stops entering and sells every open position (the kill switch); three failed transactions in a row trip a circuit breaker; a stale market feed halts entries; if live fills trail the paper mirror by more than 2 points per trade over 50 trades, entries pause and it tells you.
• Watch it think. The Overview page draws the fly's brain in three dimensions: every neuron lit by the minute's scoring, and the synapses it has changed by learning drawn as lines. Drag to orbit, hover a neuron for its cell type.

Requirements

• Docker (Desktop on macOS/Windows, or Engine + Compose v2 on Linux). ~2 GB of disk for the image and database.
• A Helius API key (free tier is enough) and a Jupiter API key.
• 0.5 SOL you are prepared to lose. Not more: the caps below are sized for 0.5.
• Any 64-bit CPU: the fly scores about as fast on 16 CPU threads as on a GPU (its forward pass is a memory-bound scatter over a million synapses). An NVIDIA GPU is used when you start it with the CUDA override (see below). Apple GPUs are not reachable from inside Docker.

Start trading

1. Download this branch and open a terminal in it:

   git clone -b distribution https://github.com/bryceweiner/fly-trader.git && cd fly-trader
   

2. Create your .env and paste the two API keys into it:

   cp .env.example .env
   

3. Start it:

   docker compose up -d --build
   

   The first start builds the image (several minutes), creates the database, installs the models and starts the market feed and the trading engine.
4. Create the bot wallet — it prints the public key and writes the secret into your .env:

   docker compose exec fly fly-trader wallet new
   

   Back up .env now. It is the only copy of that key.
5. Send 0.5 SOL to the printed address, then restart the engine so it picks up the key:

   docker compose restart fly
   

6. Open the console: http://localhost:8501. Watch the Overview page.

That is all. From here it runs on its own. The first ~24 hours are a warm-up (it needs a day of candles before it scores anything); then it trades paper until 20 trades have closed; then it trades the wallet.

Day to day

The console listens on 127.0.0.1 only. Do not expose it: it can start and stop the engine and clear the rails.

What is in the box

The fly's replay verdict shipped here: over the 28 evaluation days it made +1.52 % per trade after real costs (profit factor 1.13) against −5.27 % for random picks of the same tokens. That is a backtest on one corpus, not a forecast.

Tuning

Edit .env, then docker compose restart fly. The ones that matter:

Scaling the wallet up scales the position caps with it (MAX_POSITION_FRACTION), but MAX_POSITION_SOL is a hard ceiling: raise it deliberately or not at all.

NVIDIA GPU

Optional. The CPU image trades fine; a GPU makes the once-a-minute scoring and the fly's learning a little faster. You need the NVIDIA driver (560 or newer) and the NVIDIA Container Toolkit on the host; then start with the CUDA override instead of the plain command in step 3:

docker compose -f docker-compose.yml -f docker-compose.cuda.yml up -d --build

The image is built with the CUDA 12.6 torch wheel (TORCH_INDEX: cu126 in docker-compose.cuda.yml; cu130 for the newest drivers) and the container is given every GPU. DEVICE=auto in .env then resolves to the GPU; the console's Model page shows which device the fly is on. Batch sizes are derived from the card's free memory, so an 8 GB card works; on a shared card set DEVICE_MEMORY_GB in .env to what the fly may use.

Updates

The container keeps itself current. Every hour it asks Hugging Face whether bryceweiner/fly-trader has a new release; when it has, it downloads the new code and model, restarts (about a minute; the paper books, the fly's learned state and any open positions are untouched -- open positions are sold on schedule by the restarted engine) and keeps the previous release beside the new one. If the new release does not come up within five minutes it goes back to the previous one and never tries that release again. docker compose logs -f fly shows what it did. AUTO_UPDATE=0 in .env turns this off; then the "Refreshing the model" steps below are the manual way.

Refreshing the model

The model is a snapshot of one training run. To ship a newer one, on a master checkout with its database, after a passed replay:

.venv/bin/python tools/make_seed.py --out dist
cp -R dist/models/* models/ && cp -R dist/seed/* seed/

Commit models/ and seed/ on this branch. The seed is idempotent: an install that already has these rows is not changed by it, so pull the branch, docker compose up -d --build, and the new model is in place.

Also on Hugging Face

• Model and this code: bryceweiner/fly-trader
• The training database (the full corpus, for retraining on master): bryceweiner/fly-trader-seed

Credits

FlyWire consortium, Dorkenwald et al., Neuronal wiring diagram of an adult brain, Nature 2024; annotations Schlegel et al. 2024; synapse signs Shiu et al. 2024. Connectome export: snedea/flybrain (MIT). Market data: PumpAPI; routing: Jupiter.

MIT.