165,122 neurons steer the cursor, while the rig finishes the launch
flycoinrh builds a graph from a male fruit-fly connectome containing 165,122 traced neurons and 10,228,000 signed connections. A Chromium screenshot is reduced through 892 retinal columns, neuron activity is simulated, and four descending-neuron groups control horizontal movement, vertical movement, reverse, and click. That loop is genuinely more interesting than a random cursor because the wiring comes from measured anatomy.
The fly does not independently launch a token. The README says it reliably fills the description, often fills the ticker, rarely fills the name, and does not reach the launch button alone. Human-authored rig code completes missing fields, picks one of 57 paired assets, sets the creator tax and social handle, and presses Confirm. The project discloses that boundary instead of dressing the whole sequence up as autonomy.
What happened when we ran it
Our sandbox cloned commit f2fdedf, a repository with 134 files, roughly 36,155 lines of source, and a 5.2 MB checkout. Python installation succeeded in 32 seconds, pulling 79 packages and using 609 MB. The build passed in 5 seconds. Those results cover repository mechanics, not a connectome build, Chromium session, wallet signature, or live chain transaction.
No test script or standard target was available to our harness, so it skipped tests. The repository puts many test_*.py files at its root, but our supplied measurement has no executed cases and no counts. Our scan also found no CI workflow and no tests directory. A buyer should run the intended root test commands and record the result before using any path that can browse, post, or sign.
Pip-audit found 37 known vulnerabilities in the installed environment. The report supplied to us does not break those findings down by package or severity, so we cannot say which are exploitable in this application. We can say that a wallet-bearing process with 37 reported advisories is not ready for funds until every item has been traced, patched or accepted with a documented boundary.
The 1.1 GB connectome is outside the checkout
The repository does not ship its largest input. The README directs users to download a 1.1 GB connectome-weight file plus 14 MB of annotations and 42 MB of neurotransmitter data, then run build_graph.py. The result should contain the 165,122 neurons and 10,228,000 edges used in the published experiment. The data remains CC-BY even though the project code is MIT.
Running the browser also requires Playwright's Chromium. The full Python dependency set includes scientific arrays, Feather readers, FastAPI, browser automation, and Ethereum signing. Our 609 MB environment is therefore only one part of local storage. The separate Docker path expects the generated graph and data to exist before build, because a deploy made straight from GitHub has nothing to load.
Two flags stand between a dry run and a transaction
FLY_ALLOW_BROWSER=1 permits a real site to open, while FLY_RH_LIVE=1 permits transaction signing. Both default to off. The documented dry run checks chain ID, nonce, gas, signing, address recovery, and estimation without broadcasting. Live use creates a secp256k1 wallet, keeps its secret in a gitignored environment file, injects an EIP-1193 provider into the launchpad, and signs the site's request in Python.
This is still cryptocurrency automation. The README records a live $FLYBRAIN launch and gives its transaction and contract, but it also tells readers not to buy the test tokens. The project is strongest as evidence that the interface loop reached an on-chain outcome. It offers no reason to treat the resulting asset as an investment, and the fly's behavior provides no market signal.
The roaming container has no wallet by design
roam.py gives the connectome a browser with no instructions, but the public service removes the wallet, keyboard, downloads, popups, dialogs, and sensitive click targets. Its first keyword blocklist failed, according to the README, so the current design uses an allowlist of link-rich domains. FLY_ROAM_OPEN=1 removes that fence and is explicitly unsuitable for an unattended public stream.
The Dockerfile reinforces the separation by omitting eth-account and private keys. An optional paper executor can read quotes and maintain a pretend ledger, but it cannot sign. The language-model narrator is another separate layer: it receives allowlisted observations, rejects ungrounded numbers and trading language, then may post through capped X credentials. Any first-person curiosity in those posts belongs to the narrator, not the fruit fly.
Eight open issues and PRs expose useful scientific disputes
The repository was created September 10, 2026 and pushed through September 14. GitHub showed 191 stars, 62 forks, and eight open issues and pull requests on October 2. One open pull request argues that the code classified reward and punishment compartments from the wrong matrix direction, and says the README's earlier controlled result needs to be rerun under the corrected split.
That dispute is exactly why the repo works better as an experiment than a finished platform. Biological data, simulator choices, visual encoding, reward definition, web selectors, and transaction code all meet in one result. A green 5-second build cannot validate that chain of reasoning. The honest next step is to isolate the part you care about, reproduce it with fresh controls, and leave the funded wallet out until the 37 audit findings and unexecuted tests are resolved.