Real-time 3D in the browser. Two players, one arena. What makes it different is that damage is not a number — it is an anatomical simulation.
A round through the skull ends you instantly. A round through the right forearm severs the flexor tendons, and your hand stops working: you cannot hold a weapon, and you cannot work a tourniquet with that hand either. A round through the front of the thigh clips the femoral artery and starts a two-minute bleed-out clock; the same round through the back of the same thigh takes only hamstring and you survive it. Nothing about that asymmetry is special-cased — the artery is simply anterior-medial, and the geometry decides.
There is no HP, no armour value, and no damage number anywhere in the codebase.
Python 3.12+ and a Chromium or Firefox browser. No Node.js, no npm, no build step. Everything is free and permissively licensed; the project costs nothing to develop or play and runs fully offline after setup.
Windows (PowerShell):
powershell -ExecutionPolicy Bypass -File tools\bootstrap_venv.ps1Linux / macOS (Bash):
chmod +x tools/bootstrap_venv.sh
./tools/bootstrap_venv.shThat creates .venv, installs four dependencies, vendors three.js (pinned and
sha256-verified), and generates a self-signed TLS certificate.
The server is never started automatically.
.venv\Scripts\python.exe run_server.py --no-tls # same PC, two windows
.venv\Scripts\python.exe run_server.py # LAN, TLS for the 2nd machine
.venv\Scripts\python.exe run_server.py --no-spin # cooler on battery
Same PC: open http://localhost:8443 in two windows. No certificate needed —
http://localhost already counts as a secure context, so mouse capture works.
Over LAN: run with TLS and open https://<your-lan-ip>:8443 on the second
machine. Accept the certificate warning once. This is not optional decoration:
the Pointer Lock API requires a secure context, and a plain http:// LAN
address is not one, so mouse capture will silently fail without TLS. The
certificate script prints the correct URL.
W A S D |
move — backpedal and strafe are slower than forward |
Shift |
sprint (costs stamina; stamina loss widens your sway, not your speed) |
Ctrl / Z |
crouch / prone — both are latches, not holds |
Q / E |
lean around cover |
Space |
jump; against a ledge, mantle |
Alt (right) |
free-look — check your flank without changing course |
Right Shift |
hold breath — steadies the sights for up to 6 s, then costs you |
| Mouse 1 / 2 | fire / aim down sights |
R |
reload |
1–4 |
rifle, SMG, pistol, knife |
V |
shoulder swap |
H |
medical panel |
5–0 |
tourniquet, gauze, bandage, chest seal, splint, pill pack |
F1 |
hitbox view — draws the server's actual collision volumes |
F2 |
performance and prediction-divergence overlay |
Blood volume, per-organ involvement, named arteries, limb motor integrity, pain, shock and consciousness, all integrated continuously. Representative outcomes, all verified by the test suite:
| Wound | Outcome |
|---|---|
| Cranium, brainstem, cervical cord | instant, untreatable |
| Heart | unconscious ~5 s |
| Abdominal aorta | unconscious ~21 s |
| Carotid | unconscious ~7 s — at 80%+ blood remaining, because it starves the brain directly rather than draining you |
| Iliac artery | ~59 s, and no tourniquet reaches it |
| Femoral artery | ~112 s untreated; survivable with a tourniquet inside the first minute |
| Liver | ~70 s |
| Gut, kidney, single lung | survivable — conscious, in agony, badly degraded |
Limb consequences: right arm disabled forces the left hand (accuracy ×0.45, sway ×2.2, rifle ADS impossible). Both arms disabled means no attack at all — but movement is completely unaffected. A destroyed leg forces prone and a 0.65 m/s crawl, which is deliberately not zero: you need to be able to drag yourself to cover and get a tourniquet on, and that decision is the whole point.
Constrained to what a soldier with an IFAK can actually do. A tourniquet works on limbs and nowhere else — attempting one on the abdomen is rejected with a reason, not silently ignored, because a treatment that quietly does nothing teaches nothing. Nothing in the kit reaches the brain, spine or heart. One of each item per life, so treatment is triage.
The morphine is a trap, and a real one: it cuts pain 70% and steadies your aim while doing nothing whatsoever about the bleeding, and it costs consciousness margin you may need.
.venv\Scripts\python.exe -m pytest
217 tests. The suite covers physiological trajectories against documented realistic bands, TCCC eligibility rules, the character controller (asserting the capsule never ends a tick inside geometry), collision primitives against brute-force intersection, wire-format quantization, and a full headless integration pass driving the real world through the real codec.
Cross-language parity is the one thing pytest cannot check. movement.py
and movement.js implement the same controller in two languages, so
tools/gen_movement_golden.py records a 1275-frame trajectory from the
authoritative Python and client/tests.html replays it through the JavaScript,
reporting the worst divergence against the 10 mm reconciliation tolerance. Open
that page through the server to verify it.
.venv\Scripts\python.exe tools\check_client.py # static client wiring check
server/physio/ the wound model — pure, no I/O, no clock, no RNG unless passed one
server/sim/ trace, movement, ballistics, rig, hit registration, knife, lag comp
server/net/ wire codec, transport, sessions
shared/*.json read verbatim by BOTH sides — the anti-drift mechanism
client/js/ three.js renderer, prediction, HUD, WebAudio
Three decisions shape everything else:
Wound state is authoritative-only. It is never predicted and never replayed, so the entire physiology model carries no cross-language determinism burden and can be pure, deterministic Python that is exhaustively unit-tested.
The rendered player model is the hitbox rig. The same 22 boxes the server ray-tests are the ones three.js draws. Hitbox/visual drift — the dominant source of "that was bullshit" in shooters — is structurally impossible. Press F1 to see them.
Every shared constant lives in shared/*.json. Neither movement.py nor
movement.js may hardcode one; a test enforces it. That is what keeps
prediction and authority in agreement.
~50 MB total (37 MB of it the venv), zero binary assets — every texture is procedural, every sound is WebAudio-synthesised, all geometry is generated from JSON. Server load is around 1.6 ms per tick with both players firing, about 10% of the 15.6 ms budget at 64 Hz. Snapshots are ~110 bytes, roughly 7 kB/s per client. The renderer targets the integrated GPU at a 60 FPS cap and pauses on window blur.
This project is licensed under the MIT License.
| Component | Licence |
|---|---|
| Python 3.12 | PSF-2.0 |
| websockets | BSD-3-Clause |
| cryptography | Apache-2.0 OR BSD-3-Clause |
| pytest | MIT |
| pytest-benchmark | BSD-2-Clause |
| three.js | MIT |
All permissive. No account, API key, subscription or hosting bill, ever.
We welcome contributions! Please review our Contributing Guide and Code of Conduct before submitting pull requests or opening issues. Security concerns should follow our Security Policy.