# 🪰 Awesome Fly Brain Agents

> **Real Drosophila connectomes, actually running things.**

A curated list of projects that put large-scale fruit-fly connectomes into an active loop: games, robots, simulated bodies, markets, desktop agents, and other environments.

[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE) [![PRs Welcome](https://img.shields.io/badge/PRs-welcome-brightgreen.svg)](CONTRIBUTING.md)

## 🎮 Games

| Project | Connectome / reported scale | Environment | Closed loop | Notes |
|---|---|---|:---:|---|
| [nftechie/doomfly](https://github.com/nftechie/doomfly) | MaleCNS, 166,700 neurons | Doom / ViZDoom | ✅ | Connectome-based spiking simulation drives an engineered Doom interface. The project explicitly cautions against interpreting this as validated fly vision or demonstrated learning. |
| [ornata/fly](https://github.com/ornata/fly) | MaleCNS | Super Mario 64 | ✅ | Fly64 connects a MaleCNS-based brain model to SM64 through a visual-input → neural-network → controller loop. |
| [blendi-remade/fly-brain-minecraft](https://github.com/blendi-remade/fly-brain-minecraft) | MaleCNS, 176,422 neurons | Minecraft | ✅ | Minecraft sensory events drive real sensory populations; descending and motor activity controls a fly mob. |
| [michaelpersonal/flytype](https://github.com/michaelpersonal/flytype) | MaleCNS, 166,700 neurons | Typing + brick breaker | ✅ | Rendered pixels drive the connectome and neural activity is read back into actions. |
| [MidTermDev/immortal-fruit-fly](https://github.com/MidTermDev/immortal-fruit-fly) | FlyWire, 139,248 neurons | Arena + Doom | ✅ | Whole-brain LIF simulation with embodied arena and Doom modes; also records state hashes on-chain. |
| [dohun1214/malecns-asteroids](https://github.com/dohun1214/malecns-asteroids) | MaleCNS, 166,700 neurons | Atari Asteroids | ✅ | Asteroid positions stimulate visual populations in a whole-brain LIF model; motor readouts steer escape or pursuit. Includes lesion and rewiring controls; firing and control timing use engineered rules. |
| [seanphan/flyt3](https://github.com/seanphan/flyt3) | MaleCNS, 166,700 neurons | Tic-tac-toe | ✅ | Board state drives a frozen LIF connectome and a trained motor-spike readout selects moves. The play server also applies hand-written win/block overrides; gameplay is not solely a neural readout. |
| [Reldnahc/pokefly](https://github.com/Reldnahc/pokefly) | MaleCNS, repo reports 166,700 neurons | Pokemon Red / PyBoy | ✅ | Screen pixels drive a connectome simulation with fixed motor-to-button mappings and experimental internal synaptic plasticity. Outcome rewards use game telemetry; useful screen-specific gameplay learning remains unestablished. |
| [shantanugoel/fly-games](https://github.com/shantanugoel/fly-games) | MaleCNS, repo reports 166,700 neurons | Super Mario Bros. / Kung Fu / Doom | ✅ | Structured game facts stimulate a frozen spiking connectome; descending activity feeds a fitted readout or hand-written decoder. Only the readout is trained, and held-out imitation scores do not establish closed-loop game success. |
| [almera-vs/malecns-pong-lab](https://github.com/almera-vs/malecns-pong-lab) | MaleCNS, repo reports 166,700 neurons | Pong with simulated limb-and-paddle control | ✅ | Pooled court pixels and joint feedback drive sparse spiking activity; tibia motor pools control engineered paddles. A three-seed pilot changed synaptic weights without improving held-out hit/miss counts. |
| [HakimElAyoubi/fly-chess](https://github.com/HakimElAyoubi/fly-chess) | MaleCNS, repo reports 144,209 neurons | Chess / UCI engine | ✅ | Encoded board state drives photoreceptors in a rate-based connectome; a trained descending-neuron readout selects moves. Network gains, biases and readout heads are trained; legal-move and anti-repetition rules constrain play. The walking fly in the demo is animated. |
| [blackicon-eth/fly-plays-games](https://github.com/blackicon-eth/fly-plays-games) | MaleCNS, repo reports 166,700 neurons | Retroid / Pokemon Red via PyBoy | ✅ | Game-supplied object positions drive visual projection neurons in a frozen LIF connectome; a trained descending-neuron readout controls the Retroid paddle. Only the readout is trained; the long Pokemon journey is scripted, not neural navigation. |
| [gabinacode/fox-vs-fly](https://github.com/gabinacode/fox-vs-fly) | MaleCNS, repo reports 166,700 neurons | Browser platform fighter | ✅ | Six engineered game sensors drive a full-graph rate model; a calibrated descending/motor readout selects movement and attacks. The model uses positive synapse counts without transmitter signs, and recovery logic and action gates are authored rather than learned. |
| [CodeMan1729/flybrain-arena](https://github.com/CodeMan1729/flybrain-arena) | MaleCNS, repo reports 166,700 neurons | Godot drone combat | ✅ | Aim/fire cues enter LC4/LPLC2 directly in a full-graph signed-rate model; DNp01/DNp03 activity is decoded into evasive motion. This FLYFEAR derivative combines neural dodge output with engineered pursuit and flight rules; retinal looming detection is not simulated. |
| [jamesbiederbeck/flappy-haltere](https://github.com/jamesbiederbeck/flappy-haltere) | MaleCNS v1.0, full-graph LIF model | Flappy Bird | ✅ | Pixels and fall-speed-derived haltere stimulation drive a shared connectome simulator; accumulated flight-muscle motor spikes trigger flaps. Input mappings and the flap threshold are engineered; reported play mainly holds altitude, without demonstrated active pipe avoidance. The engine requires a separate checkout and local graph preparation. |
| [icybb0903-sketch/flygo](https://github.com/icybb0903-sketch/flygo) | MaleCNS, repo reports 139,662 nodes / 5,536,347 connections | 9×9 Go | ✅ | Encoded board state drives a fixed-graph LIF simulation; pooled neural features feed an offline-trained linear readout with legal-move masking. Input/output pools are engineered, play does not train the policy online, and playing strength remains unestablished; the fly carrying stones is animated. |
| [professorwang/flybrain-banana-quest](https://github.com/professorwang/flybrain-banana-quest) | Repo reports FlyWire 139,255 / MaleCNS 176,422 neurons | Browser banana-foraging arena | ✅ | Modeled odor and taste drive a full-graph LIF network; descending activity steers movement and feeding-pool spikes gate consumption. Sensory pools, weight normalization and readouts are engineered; low-activity random wandering, food-proximity stopping and wall bounces are game rules, with no synaptic learning. |
| [0Sakura721/flyaim](https://github.com/0Sakura721/flyaim) | MaleCNS, repo reports 166,700 neurons | Offline aiming arena | ✅ | Pixels drive a full-brain LIF model; a trained descending-neuron readout moves the crosshair. The reported frozen-connectome experiment found no advantage over shuffled wiring and worse target distance than random actions; retinal mapping and transmitter signs include known approximations. |
| [castor639/fly-tennis](https://github.com/castor639/fly-tennis) | MaleCNS, repo reports 163,903 neurons | Simulated racket game | ✅ | Analytic compound-eye input drives full-graph rate networks; medulla activity relative to a ball-free twin network controls lateral positioning. Edges below five synapses are removed; the untrained match uses engineered decoding, while swings, return targets and recentering are scripted and legs are not simulated. |
| [ThomasLin070217/FlyGift](https://github.com/ThomasLin070217/FlyGift) | FlyWire v783, repo reports 139,255 neurons | macOS fly-swatting game | ✅ | Swatter approach and body feedback drive a whole-brain Metal LIF simulation; Giant Fiber spikes trigger escape and descending activity modulates movement. This SiliconFly derivative uses engineered sensory gains, procedural gait/flight, target selection and game timing; it has no synaptic learning or simulated ventral nerve cord. |
| [Zaib111/fly-clash-royale](https://github.com/Zaib111/fly-clash-royale) | FlyWire v783, repo reports about 134k neurons / 2.7M connections | Clash Royale clone | ✅ | Structured arena and card-state inputs drive a whole-brain rate model; a descending/motor readout selects cards and placements. Edges below five synapses and isolated neurons are removed; PPO self-play trains input/output mappings, biases and gains while retaining topology and signs. Legal-action masking is engineered; playing strength and a benefit from fly wiring are not established. |
| [rustyorb/barfly](https://github.com/rustyorb/barfly) | FlyWire v783, repo reports 138,639 neurons / 15.1M connections | Interactive browser bar scene | ✅ | Sugar and looming stimulation drive a whole-brain LIF model; proboscis and Giant Fiber activity control drinking and flinching. Drinking depletes the mug and raises an engineered GABA-gain feedback signal; sensory mappings, animation and intoxication rules are authored, with no training. Dialogue comes from a separate language model and does not control the connectome. |
| [Arkazzae/FlyChess](https://github.com/Arkazzae/FlyChess) | FlyWire v783, repo reports 134,181 neurons / 2.7M connections | Browser chess | ✅ | Encoded board features drive ten steps of a whole-brain rate model; trained policy/value readouts guide PUCT search. Connections require at least five synapses; topology stays fixed while gains, biases and interfaces learn from chess labels including Stockfish analysis. Rules and search are engineered; the reported rating is specific to its limited evaluation protocol, not a FIDE rating. |
| [seruyvodoley/flyswarm-connectome](https://github.com/seruyvodoley/flyswarm-connectome) | MaleCNS; repo reports 2,667,200 neural states across 16 agents | Godot tank combat | ✅ | Structured visual cues and optional inter-agent audio drive full-CNS FlyBrain spiking models; descending activity feeds a fixed decoder or trained external motor adapter. The connectome stays fixed. Navigation, objectives, aiming and firing also use explicit assistance rules; separate Rule AI and MARL modes do not demonstrate connectome control. |
| [dorywoof/BioGrid](https://github.com/dorywoof/BioGrid) | FlyWire v783, repo reports 138,639 neurons | Water- and shade-seeking flight game | ✅ | Modeled humidity and cooling stimulate sensory populations in a whole-brain LIF model; left/right DNb05 firing differences steer flight. Kenyon-cell connections, inputs onto sensory neurons and monoaminergic outputs are removed. Sensory encoding, steering interpretation, spontaneous turns, landing and physiology are engineered; reported lesion effects are simulation results, not experimental validation in flies. |
| [OrneyKraken/fastfly](https://github.com/OrneyKraken/fastfly) | FlyWire v783, repo reports 138,639 neurons / 15.1M connections | User-drawn 2D mazes | ✅ | Goal bearing and wall proximity drive LC10a/LPLC2 populations in a whole-brain LIF model; calibrated DNa01/DNa02 firing differences steer movement. Simulation retains connections with at least one synapse and uses fixed weights. Sensory encoding, readout normalization, speed, random wandering and wall sliding are engineered; reported maze runs do not establish learned navigation. |
| [BuruCebe/fly-brain-maze](https://github.com/BuruCebe/fly-brain-maze) | FlyWire v783, repo reports 138,639 neurons / 15M connections in Python mode | Food-seeking mazes, solo or packs | ✅ | Modeled odor drives a whole-brain LIF network with added antennal-lobe synaptic depression; a fitted 40-neuron descending readout steers walking. Smell contrast, fixed speed, random wandering and wall turns are engineered. The browser version uses a trimmed graph, and its optional food-mark attraction adds a direct steering rule; reported group benefits are not evidence of emergent cooperation. |
| [S1MS4/fly-chess](https://github.com/S1MS4/fly-chess) | MaleCNS v1.0, repo reports 163,803 neurons / 6,234,868 connections | Browser chess | ✅ | Board planes drive 12 steps of a full-graph rate model; trained descending-neuron readouts score legal moves. The backprop model trains edge magnitudes, neuron gains/biases and output heads; connections below five synapses, self-edges and isolated nodes are removed, and photoreceptor outputs are made excitatory. Repetition avoidance and optional one-ply evaluation are engineered; the repo reports shuffled wiring outperforming the biological graph. |

## 🤖 Robots, bodies & control

| Project | Connectome / reported scale | Environment | Closed loop | Notes |
|---|---|---|:---:|---|
| [FutureJJ/ommatid](https://github.com/FutureJJ/ommatid) | MaleCNS, 165,122 CNS neurons + FlyVis visual frontend | Physical hexapod | ✅ | Camera input and descending-neuron readouts connect to a physical hexapod. Direct leg-motor control is in dry run; tested reflexes were not established. |
| [MarkUnthank/flyhard](https://github.com/MarkUnthank/flyhard) | MaleCNS, 165,122 traced neurons | Simulated body + wheel + CARLA | 🧪 | Reports trained steering through a simulated foreleg and wheel connected to CARLA. Turns are instructed and speed is scripted; visual driving remains untested. |
| [rembish/fruit-fly](https://github.com/rembish/fruit-fly) | FlyWire, 139,255 neurons | Desktop sprite | ✅ | A desktop fly driven by a whole-brain LIF simulation; cursor / looming input is mapped into identified circuits. |
| [nsfm/fly-afterlife](https://github.com/nsfm/fly-afterlife) | MaleCNS, 162,517 neurons + FlyVis frontend | Simulated room | ✅ | Visual and contact inputs drive a spiking connectome; descending and leg-motor readouts control heading and pace. Loom-selectivity claims were withdrawn, and steering robustness varies across visual models. |
| [PtPavloTkachenko/fly-brain-spectacles](https://github.com/PtPavloTkachenko/fly-brain-spectacles) | MaleCNS, repo reports 166,700 neurons | AR fly on Snap Spectacles | ✅ | A browser-based LIF simulation receives room and headset cues; descending and motor readouts control a virtual fly. Visual features, walking rhythm and flight/landing control are engineered; pixels alone do not drive behavior in this model. |
| [Noir-infini/pianist-fly](https://github.com/Noir-infini/pianist-fly) | MaleCNS, repo reports 166,700 neurons | MuJoCo fly + piano | 🧪 | Modeled sugar and visual inputs drive a fixed-weight LIF connectome; motor-pool rates modulate scripted key-press timing. The score, target keys and leg trajectories are programmed; the directional gate reads plume data, and zero motor activity falls back to default timing. |
| [dimm0909/fly-in-the-bottle](https://github.com/dimm0909/fly-in-the-bottle) | MaleCNS, repo reports 164,606 neurons | Desktop fly in a simulated jar | ✅ | Touch, cursor cues and body feedback drive an adaptive-threshold LIF model; motor and descending pools control motion. Connections below five synapses and motor-neuron outputs are removed; sensory mappings, gait rhythms and body physics are engineered. Without the separately installed brain, the widget uses scripted behavior. |
| [Ryans-sS/malecns-embodied-fly](https://github.com/Ryans-sS/malecns-embodied-fly) | MaleCNS, repo reports 166,700 neurons | Simulated foraging arena | ✅ | Modeled odor, taste and body feedback drive a LIF connectome; descending/leg activity controls motion and proboscis activity gates ingestion. Edges below two synapses are removed; metabolism, exploration and feeding-stop rules are engineered, with simplified dopamine-gated KC→MBON plasticity. The separate NeuroMechFly demo uses scripted gait and wandering. |
| [yusenthebot/FlyArm](https://github.com/yusenthebot/FlyArm) | MaleCNS, repo reports 166,700 neurons / 10.5M connections | MuJoCo Franka arm | ✅ | Simulator state and task cues enter a trained encoder; a frozen whole-CNS rate model feeds a trained motor decoder. Self-connections and edges below three synaptic contacts are removed; this is simulation using privileged state, with no established advantage over alternative wiring. |
| [joshuabradley012/brainfly](https://github.com/joshuabradley012/brainfly) | MaleCNS, repo reports 166,700 neurons + FlyVis frontend | NeuroMechFly / MuJoCo rotating-drum arena | ✅ | A modeled visual scene drives a whole-graph spiking/graded network; baseline-corrected DNa02 activity steers the body. The documented loop uses the project's unvalidated inherited model and an engineered descending-to-gait mapping; FlyGym's walking controller supplies leg coordination. |
| [hsilvosa/fruit-fly-rl](https://github.com/hsilvosa/fruit-fly-rl) | MaleCNS, repo reports 167,184 neurons | Procedural 3D navigation | ✅ | Simulated sensors drive a fixed full-graph rate model whose features feed a learned flight policy. Sensors and flight dynamics are engineered; the current large-room demo uses explicit map planning, with a neural reader that cancels most recurrence, and does not establish learned navigation or a benefit from fly wiring. |
| [IrohAmca/NeuroPest](https://github.com/IrohAmca/NeuroPest) | FlyWire v783, repo reports 138,639 neurons in optional full mode | Desktop fly pet | ✅ | Cursor and body cues drive a LIF connectome; descending activity controls walking, retreat, escape and steering. The default uses a 15,000-neuron subset; full mode requires selecting the complete graph. Sensory mappings, state transitions and sprite motion are engineered, and the separate learning module uses synthetic mushroom-body wiring by default. |
| [thepuneeth/Flycraft](https://github.com/thepuneeth/Flycraft) | Repo reports 138,639 FlyWire brain neurons + 11,000 MaleCNS nerve-cord nodes | MuJoCo hexapod | ✅ | The brain-coupled path feeds target-bearing cues through a frozen whole-brain rate model, then a nerve-cord model and fixed motor map drive joint torques. Sensory encoding, an alignment gate and body control are engineered; reported corrected trials did not produce a coherent tripod gait or outperform matched random descending drive. The main nerve-cord-only experiments use the smaller graph. |
| [liudicsu/3dfly](https://github.com/liudicsu/3dfly) | MaleCNS v1.0; repo reports about 185k connected annotated neurons in optional full mode | MuJoCo flight and point-cloud mapping | ✅ | Camera features and body feedback drive a full-graph rate model; descending activity modulates wing strokes and body torques. The full build defaults to at least three synapses per connection and excludes unannotated or isolated nodes; the quick-start demo uses 500 synthetic neurons. Sensory encoding, readout calibration, exploration bias, hover/altitude control and collision recovery are engineered, with no validated biological flight or depth perception. |
| [kkokosa/dotFly](https://github.com/kkokosa/dotFly) | MaleCNS v1.0, repo reports 166,700 neurons | Godot living-room fly demo | ✅ | Visual, odor, wind and contact cues drive a whole-CNS LIF model; neural readouts modulate turning, speed and feeding. The demo retains superclass-annotated bodies, drops unclear/histamine outputs and silences Kenyon-cell and antennal-lobe local-neuron outputs for stability. Sensory encoding, navigation decisions, landing and body reflexes are engineered; the fixed graph does not learn, and neural navigation is not established. |
| [izntariq/fly-with-me](https://github.com/izntariq/fly-with-me) | FlyWire v783, repo reports 138,639 neurons / 15,091,983 connections per fly | Two-deck audio mixer and synthesizer | ✅ | Audio features stimulate whole-brain LIF models; descending/motor spikes drive synth events, mixing moves and deck preference, with optional neural track/exit choices. The default uses two listener flies and a DJ fly, and loads 9,978 trained synaptic weights when available. Frequency mappings, burst thresholds, habituation, beat alignment and DJ routines are engineered; reported training and neural-decision comparisons show no measurable benefit. |
| [RustamKuramshin/fly-boot](https://github.com/RustamKuramshin/fly-boot) | FlyWire v783, repo reports 138,639 neurons / 15.1M connections | HTTP request routing | 🧪 | Encoded request methods and paths stimulate a whole-brain LIF model; a trained Kenyon/projection-neuron readout selects handlers or rejects requests. Odor encoding, inhibitory gain and training examples are engineered, and paths can be misclassified; optional response-status feedback updates the readout and is disabled by default. |
| [cora1022/Flies](https://github.com/cora1022/Flies) | FlyWire v783, repo reports 138,639 neurons / 15.09M connections | FlyGym / MuJoCo wind-and-food arena | ✅ | Modeled wind and taste drive a whole-brain LIF network; calibrated descending activity controls walking and turning, while MN9 activity gates feeding. Sensory encoding, decoding, feeding timing and FlyGym leg coordination are engineered; the fixed brain graph omits the nerve cord, and reported stimulus-induced runaway activity limits interpretation. |
| [william-aqn/malecns-warhammer](https://github.com/william-aqn/malecns-warhammer) | MaleCNS v1.0, repo reports 166,700 neurons / 25,582,938 connections in full mode | Browser sensory arena | 🧪 | Sensory and body cues drive a full-graph rate model; six VNC motor-group readouts move a kinematic body when released from its default clamp. The full mode retains annotated non-glial neurons and their induced edges, with engineered sensory tuning, transmitter effects and body rules, no training and incomplete sensory coverage; reported visual motor effects are weak. The separate tactical mode uses a smaller graph. |
| [S1MS4/fly-hit-vape-and-scroll-tiktok](https://github.com/S1MS4/fly-hit-vape-and-scroll-tiktok) | MaleCNS v1.0, repo reports 164,587 neurons / 24,914,841 connections | Simulated room + TikTok feed | ✅ | Video pixels and modeled odor drive a whole-CNS LIF network; thresholded neural readouts trigger simulated puffs and real feed swipes. Self-edges and fast monoaminergic outputs are removed, photoreceptor outputs are made excitatory, and short-term depression is added. Sensory mapping, nicotine effects, action thresholds and animation are engineered, with no trained wiring; headless runs also force timeout swipes, and the repo reports likes/follows/reposts do not persist. |
| [Alexing-uni/clockwork-fly](https://github.com/Alexing-uni/clockwork-fly) | FlyWire v783, repo reports about 139k neurons / 15M connections | Video-viewing animated fly | 🧪 | Video pixels and taste inputs drive a whole-brain LIF model; descending and MN9 readouts trigger animated escape, turning and feeding responses. The body does not control video playback; sensory mapping, event thresholds, boosted visual weights and added spike-frequency adaptation are engineered. Full mode requires downloaded data, otherwise the app uses a synthetic network; displayed thoughts and distress labels are authored interpretations, not evidence of cognition or suffering. |
| [FredBourouliou/fly-seance](https://github.com/FredBourouliou/fly-seance) | FlyWire v630, whole-brain LIF model | Interactive question-and-offering scene | 🧪 | A local Gemma model converts questions into sensory stimulation; neural feeding/grooming readouts drive a 3D fly and a threshold-based verdict that Gemma verbalizes. Sensory rates, verdict thresholds and animation are engineered, with no learning or body feedback; the fly model does not understand questions, and the repo reports that the strongest offering predicts nearly all recorded verdicts. |

## 📈 Markets & trading

> These are neural-interface experiments, **not evidence of profitable trading** and not financial advice.

| Project | Connectome / reported scale | Environment | Closed loop | Notes |
|---|---|---|:---:|---|
| [nftechie/stonkfly](https://github.com/nftechie/stonkfly) | MaleCNS, 166,700 neurons / 25.6M connections | Coinbase market data + guarded trading actions | ✅ | Market state becomes sensory input; neural activity proposes buy / sell / hold. The project explicitly states that profitable learning has not been demonstrated. |
| [jthomason/fruit-fly-ipo](https://github.com/jthomason/fruit-fly-ipo) | FlyWire v630, repo reports 127,400 neurons / 14.7M connections | Monte Carlo IPO trading simulation | 🧪 | Dictionary-tagged text and engineered valuation/earnings signals stimulate sugar/bitter receptors in a whole-brain LIF model; MN9 minus DNge067 spikes determine simulated share trades. The fixed network has no learning, and trades do not feed back into neural input or market prices; long-only cash constraints and the valuation rule are authored. Reported model-market results are not a historical backtest or evidence of a trading edge. |

## Scope

### Required for the main list

- Use **measured Drosophila connectome data** such as MaleCNS or FlyWire.
- Run at **whole-brain / whole-CNS or comparable large scale** — generally around 100k+ neurons.
- Actually **simulate or propagate neural activity**, rather than only visualize the graph.
- Feed sensory or external information into the neural system.
- Read neural activity back out as actions, controls, or decisions.
- Provide public code, technical documentation, or enough evidence to inspect what is happening.

A **closed sensor → brain → action → environment loop** is preferred. Partial control experiments must clearly state what has been demonstrated and what remains untested.

### Usually not in the main list

- Static connectome viewers.
- Small hand-selected circuits.
- Generic neural networks merely *inspired by* flies.
- Compound-eye or insect-inspired robotics that do not use measured connectome data.
- Projects with only a viral video and no inspectable technical material.

Only existing projects supported by inspectable evidence belong in the main list. Do not add wanted entries or placeholder categories.

## Important caveat

A connectome is a wiring diagram, not a complete biological brain model.

Every runnable project must make additional assumptions about some combination of:

- neuron dynamics,
- synaptic signs and weights,
- membrane parameters,
- sensory encoding,
- motor decoding,
- plasticity,
- reward signals,
- time scaling,
- missing or uncertain biological data.

> **Real connectome ≠ reconstructed mind.**<br>**Controls a game ≠ understands the game.**

Neuron counts also vary between projects because filtering and inclusion policies differ. Counts in the list are reported according to each project's public documentation and should not be assumed to be directly comparable.

## Reading the tables

- **Closed loop** — environment affects neural input and neural output affects the next environment state.
- **✅** — a closed loop is documented; this does not establish successful learning or biological fidelity.
- **🧪** — partial or experimental integration; see the entry's limitations.

The original nine project descriptions were checked against public repository documentation on **2026-09-16**. Entries for `dohun1214/malecns-asteroids`, `seanphan/flyt3`, and `nsfm/fly-afterlife` were checked against documentation and control-loop source on **2026-09-17**. `Reldnahc/pokefly` was checked against documentation and control-loop source on **2026-09-18**. `shantanugoel/fly-games` and `almera-vs/malecns-pong-lab` were checked against documentation and control-loop source on **2026-09-19**. `HakimElAyoubi/fly-chess` was checked against documentation and control-loop source on **2026-09-20**. `blackicon-eth/fly-plays-games` was checked against documentation, the Retroid control loop and upstream neural source on **2026-09-21**. `PtPavloTkachenko/fly-brain-spectacles` was checked against documentation and neural/control-loop source on **2026-09-22**. `Noir-infini/pianist-fly` was checked against documentation and neural/control-loop source on **2026-09-23**. `gabinacode/fox-vs-fly` and `CodeMan1729/flybrain-arena` were checked against documentation and neural/control-loop source on **2026-09-24**. `jamesbiederbeck/flappy-haltere` was checked against documentation, control-loop source and its shared neural engine on **2026-09-25**. `dimm0909/fly-in-the-bottle` was checked against documentation, graph construction and neural/body-loop source on **2026-09-26**. `icybb0903-sketch/flygo` was checked against documentation, neural/control-loop source and default readout metadata on **2026-09-28**. `professorwang/flybrain-banana-quest` was checked against documentation and neural/game-loop source on **2026-09-30**. `Ryans-sS/malecns-embodied-fly` was checked against documentation, graph construction and neural/arena-loop source on **2026-10-02**. `0Sakura721/flyaim` and `yusenthebot/FlyArm` were checked against documentation, graph metadata/construction and neural/control-loop source on **2026-10-03**. `joshuabradley012/brainfly` was checked against documentation and neural/body-loop source on **2026-10-04**. `castor639/fly-tennis` and `hsilvosa/fruit-fly-rl` were checked against documentation, graph construction/model details and neural/control-loop source on **2026-10-05**. `ThomasLin070217/FlyGift` was checked against documentation, packaged graph metadata and neural/game-loop source on **2026-10-07**. `IrohAmca/NeuroPest` was checked against documentation, full-graph selection/construction and neural/desktop-loop source on **2026-10-09**. `Zaib111/fly-clash-royale` was checked against documentation, graph construction and neural/game-loop and training source on **2026-10-10**. `rustyorb/barfly` was checked against documentation, graph construction and neural/scene-loop source on **2026-10-12**. `Arkazzae/FlyChess` was checked against documentation, neural inference and search integration source on **2026-10-13**. `thepuneeth/Flycraft` was checked against documentation, methods and the full-brain/hexapod control-loop source on **2026-10-14**. `seruyvodoley/flyswarm-connectome` and `dorywoof/BioGrid` were checked against documentation, neural engines, graph construction/filtering and game-loop source on **2026-10-15**. `OrneyKraken/fastfly` was checked against documentation, graph loading and neural/maze-loop source on **2026-10-17**. `liudicsu/3dfly` was checked against documentation, full-graph construction and neural/MuJoCo-loop source on **2026-10-17**. `BuruCebe/fly-brain-maze` was checked against documentation, full Python neural/maze-loop source and browser graph filtering on **2026-10-18**. `kkokosa/dotFly` was checked against documentation, neural engine and sensory/readout/body-loop source on **2026-10-19**. `izntariq/fly-with-me` was checked against documentation, graph loading, neural engine and audio/control-loop source on **2026-10-22**. `S1MS4/fly-chess` was checked against documentation, graph construction, rate-model and move-selection source on **2026-10-22**. `RustamKuramshin/fly-boot` was checked against documentation, graph conversion, neural simulation and HTTP dispatch/feedback source on **2026-10-24**. `cora1022/Flies` was checked against documentation, graph loading, neural simulation and body-loop source on **2026-10-25**. `william-aqn/malecns-warhammer` was checked against full-mode documentation, graph construction, neural propagation and sensory/body-loop source on **2026-10-26**. `S1MS4/fly-hit-vape-and-scroll-tiktok` was checked against documentation, graph construction, neural simulation and sensory/action/feed-loop source on **2026-10-27**. `jthomason/fruit-fly-ipo` was checked against documentation, graph loading, neural simulation and trading-loop source on **2026-10-29**. `Alexing-uni/clockwork-fly` was checked against documentation, full-graph loading, neural simulation, event decoding and animated-body source on **2026-10-30**. `FredBourouliou/fly-seance` was checked against documentation, full-graph simulation, sensory/readout and animated-body integration source on **2026-10-31**. These are source reviews, not independent reproductions.

## 🧪 Inclusion checklist

Before adding a project, try to answer:

1. Which connectome is used?
2. How many neurons are actually simulated?
3. What edge / synapse filtering policy is used?
4. What neuronal dynamics are used?
5. What exactly enters the network?
6. Which neurons or readouts determine actions?
7. Is the environment genuinely closed-loop?
8. Is anything trained or optimized?
9. If so, which parameters are trained?
10. Can another person inspect or reproduce the setup?

When a number or capability comes only from a project README, describe it as **the project's reported value or claim**.

## 🧭 Related lists & resources

This repository intentionally overlaps with broader Drosophila collections while keeping a narrower inclusion rule.

- [cobanov/awesome-fly](https://github.com/cobanov/awesome-fly) — broad collection of fruit-fly connectome projects, tools, datasets, games, and resources.
- [watthem/awesome-fruit-fly-connectome](https://github.com/watthem/awesome-fruit-fly-connectome) — broad connectome resource list covering datasets, papers, tools, analysis libraries, simulations, and community experiments.
- [natverse/malecns](https://github.com/natverse/malecns) — programmatic access and utilities for MaleCNS data.

## 🤝 Contributing

Found a real fly connectome controlling something strange, useful, or delightful?

Read [CONTRIBUTING.md](CONTRIBUTING.md), then [suggest a project](https://github.com/PerturbationAI/awesome-fly-brain-agents/issues/new?template=suggest-project.yml), [submit a correction](https://github.com/PerturbationAI/awesome-fly-brain-agents/issues/new?template=correction.yml), or open a pull request using the [PR checklist](.github/pull_request_template.md).

## License

This curated list and original repository text are released under the [MIT License](LICENSE).

Individual linked projects, datasets, papers, images, and other third-party materials retain their own licenses and terms.

**🪰 One connectome. 100,000+ neurons. Give it something to do.**