FlyWire Coding Cortex

Pip-installable connector for AI coding agents: install once, then every Cursor / OpenClaw / Claude session can run FlyWire-derived neurons (leaky integrate-and-fire on real synapse edges) and map population rates into coding drives (implement, stop, reverse, steer, groom).

Honesty bar: Wiring and contact counts come from public FlyWire data. Membrane dynamics, task→current encoding, and the rate→coding map are models. This does not claim biologically calibrated coding skill or that a fruit fly “understands” your codebase.

Why this exists

Most agent “skills” are markdown instructions. This package is a pip-installable tool plus an ambient skill/MCP connector so agents can call a real runtime every session—without cloning a folder into every project.

coding task
    │
    ▼
sense (risk / tests / urgency → currents)
    │
    ▼
LIF on FlyWire-derived circuit.json
    │
    ▼
population rates (gf, dnp09, mdn, loom, …)
    │
    ▼
CodingSignals (primary drive + clamps)
    │
    ▼
agent acts  ──verify──►  Hebbian coding memory

Quick start

# From this repository
pip install -e .
# Optional (needed for --profile full):
# pip install -e ".[full]"

flywire-cortex fetch --profile seed
flywire-cortex install
flywire-cortex status

Smoke test:

flywire-cortex reset
flywire-cortex sense --text "urgent failing auth bug"
flywire-cortex step --ms 50
flywire-cortex signals

Example: giant-fiber stop drive

flywire-cortex reset
flywire-cortex stimulate gf --strength 0.5 --ms 80
flywire-cortex step --ms 10
flywire-cortex signals
# → primary: "gf", escape: true, stop: true

If pip install fails (SSL / corporate proxy), you can still run from source:

$env:PYTHONPATH = "C:\path\to\flywire-coding-cortex\src"
python -m flywire_coding_cortex.cli status

Install

1. Package

pip install flywire-coding-cortex          # when published to PyPI
# or
pip install -e .                          # from a clone
# or
uv tool install .

pip install does not download connectome dumps. Data is opt-in via fetch.

2. Circuit profile

flywire-cortex fetch --list-profiles
flywire-cortex fetch --profile seed       # default, offline after copy

3. Wire agents

flywire-cortex install
# or subset:
flywire-cortex install --targets cursor openclaw

Copies skill/ into:

Also writes ~/.flywire-coding-cortex/mcp.snippet.json.

4. Optional MCP connector (Cursor)

Add to your Cursor MCP config:

{
  "mcpServers": {
    "flywire-coding-cortex": {
      "command": "flywire-cortex",
      "args": ["mcp"]
    }
  }
}

Then restart the agent session. Tools exposed:

Download profiles

Choose (and later change) how much FlyWire data to pull. Profiles are cached side-by-side under ~/.flywire-coding-cortex/profiles/<name>/. status / step always use the active profile.

flywire-cortex fetch --list-profiles
flywire-cortex fetch --profile seed --force     # re-copy / switch back
flywire-cortex fetch --profile full --yes
flywire-cortex fetch --profile full --yes --include-synapses   # also ~9.5 GB synapse table
flywire-cortex status

Rules:

• Agents must never auto-run full. Ask the human first.
• full requires --yes and pip install 'flywire-coding-cortex[full]' (pandas + pyarrow).
• URLs live in data/manifest.json so mirrors can be edited without code changes.

Codex dump base (FAFB v783):

https://storage.googleapis.com/flywire-data/codex/data/fafb/783

Every coding turn (agent loop)

The bundled skill instructs agents to run this loop (CLI or MCP):

1. status — ensure a circuit exists (fetch --profile seed if missing)
2. sense --text "…" — encode the task into currents (risk, test-red, urgency, ambiguity)
3. step --ms 50 — run LIF on real edges
4. signals — read primary drive and clamps
5. Act only under that primary drive (see Roles)
6. Apply practices
7. After verify: remember strengthen|weaken

LIF voltage/rates persist across separate CLI invocations in ~/.flywire-coding-cortex/state/lif_state.json. Use flywire-cortex reset to clear between unrelated tasks.

CLI reference

flywire-cortex <command> …

remember examples

flywire-cortex remember query --text "auth middleware"
flywire-cortex remember strengthen --pre concept:tests --post role:dnp09 --why "ci green"
flywire-cortex remember weaken --pre concept:migration --post role:dnp09 --why "broke prod"
flywire-cortex remember add-node --pre concept:oauth --label "OAuth flow"
flywire-cortex remember add-edge --pre concept:oauth --post role:lc4 --why "auth is high risk"

Roles → coding drives

FlyWire-style role names are coding lenses. Rates come from the live sim; the bridge picks a winner-take-most primary.

Example signals payload:

{
  "primary": "dnp09",
  "escape": false,
  "stop": false,
  "reverse": false,
  "walk_drive": 0.42,
  "nervous": 0.12,
  "support": ["lc4"],
  "rationale": "primary=dnp09 walk=0.42 nervous=0.12 mdn=0.0 gf=False"
}

Thinking practices

Applied on top of rates (see skill/practices.md):

1. Winner-take-most — one primary drive per turn
2. Inhibition delay — high risk → list blast radius before edits
3. Recurrent working set — ≤5 active files/APIs
4. Hebbian writeback — strengthen only after verify; weaken on failure
5. Sparse coding — smallest change that satisfies the drive
6. Bilateral steer — design forks get an explicit L/R paragraph
7. Proprioception first — read neighbors before new patterns
8. Refractory — after gf / mdn, no new feature drive that turn
9. Population scores — prefer CLI sense scores over vibes
10. Consolidate — end of session: 3 new + 1 weakened memory edges

Architecture

flywire-coding-cortex/
  pyproject.toml
  README.md                 ← you are here
  INSTALL.md
  LICENSE                   ← MIT (code)
  data/
    circuit.seed.json       ← curated FlyWire subgraph (CC BY-NC)
    manifest.json           ← profile URLs / sizes
    DATA_LICENSE.md
  memory/
    seed-graph.json         ← starter coding-memory ontology
  skill/
    SKILL.md                ← Cursor / OpenClaw / Claude skill
    roles.md
    practices.md
  src/flywire_coding_cortex/
    cli.py                  ← entrypoint
    lif.py                  ← LIF on CSR adjacency
    sense.py                ← task → currents
    bridge.py               ← rates → CodingSignals
    memory.py               ← Hebbian graph
    paths.py                ← FLYWIRE_CORTEX_HOME helpers
    mcp_server.py
    install_skills.py
    assets/                 ← copies shipped in the wheel
    etl/
      download.py           ← fetch profiles
      build_circuit.py      ← Codex / Feather → circuit.json

Runtime home (FLYWIRE_CORTEX_HOME)

Default: ~/.flywire-coding-cortex/

Override:

# bash
export FLYWIRE_CORTEX_HOME=/path/to/cortex-home

# PowerShell
$env:FLYWIRE_CORTEX_HOME = "D:\cortex-home"

What is real vs modeled

Environment & dependencies

• Python ≥ 3.10 (3.10+ recommended; annotations use modern syntax)
• Required: numpy
• Optional [full]: pandas, pyarrow (Zenodo Feather profile)

pip install -e ".[full]"

Limitations (v1)

• Does not simulate the full ~139k-neuron brain every turn
• Does not call live Codex apps (sign-in wall); uses static public files
• Does not depend on DesktopFly or any overlay app
• full downloads are large; confirm disk and network before --yes
• MCP server is a minimal JSON-RPC subset (initialize, tools/list, tools/call)

Troubleshooting

Citation & attribution

Please cite FlyWire when using or redistributing circuit extracts:

• Dorkenwald, S. et al. Neuronal wiring diagram of an adult brain. Nature 634, 124–138 (2024). https://doi.org/10.1038/s41586-024-07558-y
• Schlegel, P. et al. Whole-brain annotation and multi-connectome cell typing of Drosophila. Nature 634, 139–152 (2024). https://doi.org/10.1038/s41586-024-07686-5

Project sites: flywire.ai · codex.flywire.ai

Code in this repository is MIT. Derived JSON and downloaded dumps remain under FlyWire’s CC BY-NC 4.0 terms — see data/DATA_LICENSE.md.

Contributing / publishing

git clone <this-repo>
cd flywire-coding-cortex
pip install -e ".[full]"
flywire-cortex fetch --profile seed
flywire-cortex install

To rebuild the shipped seed from Codex dumps:

flywire-cortex fetch --profile codex --force
# copies rebuilt circuit into data/circuit.seed.json and package assets

When ready for GitHub/PyPI: tag a release, ensure circuit.seed.json is committed, and document the CC BY-NC split in the release notes.