Vivarium

A small world that evolves on its own.

Vivarium is a living terrarium in a web page. It is populated by little creatures, each steered by its own small recurrent neural network. Nobody tells them how to behave. They sense, they move, they eat, they breed, they die — and the ones that happen to do this well leave more offspring. Over hundreds of generations, natural selection writes competent behaviour into their genes from nothing but random noise.

There is no scripted AI here. Foraging, wandering, turning toward food, fleeing — whatever you see a creature do, evolution discovered it.

Open index.html in a browser and watch.

Run it

It's a single static page with zero dependencies and no build step.

• Easiest: double-click index.html (it runs straight from file://).
• Or serve it (nicer for some browsers):
  Code

  Copy

  python -m http.server
  

  then open http://localhost:8000.

That's it. The simulation starts immediately.

What you're looking at

• Creatures are little oriented swimmers.
  • Their fill colour is a heritable hue — so you can literally see family lineages spread across the world and slowly drift apart in colour.
  • Their rim runs green → red with diet: green is a herbivore, red a carnivore, amber an omnivore.
  • Brightness tracks how well-fed they are; a snout points the way they face and flashes red when they bite.
• Green dots are plants — the base of the food web. They grow over time and cluster into patchy "meadows".
• Click any creature to open the inspector: its genome (size, diet, vision, age, generation, offspring) and a live diagram of its neural net, lit up as signals flow from senses → mind → muscles in real time.

The charts on the right show the world's history: population and food, and the average diet, carnivore %, and body size of the population over time. Watching those lines move is watching evolution happen.

The rules of the world

Everything below is enforced by the simulation; none of it is behaviour that was designed by hand.

Senses. Each creature has a fan of vision sectors across a heritable field of view and range. Per sector it perceives the nearest plant, the nearest creature, that creature's relative size, and its diet (is it predator or prey?). It also feels its own energy, speed, age, and an internal oscillator.

Mind. Those senses feed a fixed-topology recurrent neural network (the brain). The recurrence gives it short-term memory. Its three outputs are turn, thrust, and bite. The network's weights are genetic.

Body & metabolism. Size, diet, colour, and vision are genes too. Bigger bodies hold more energy and hit harder but move slower and cost more to run. Just existing burns energy; moving burns more. Run out and you die.

Eating. Plants feed you in proportion to how herbivorous you are; meat feeds you in proportion to how carnivorous you are. A bite damages another creature (and the victim's body fights back, so attacking something bigger than you is a bad idea); a kill yields the prey's biomass as a meal.

Reproduction. Asexual. Once mature and well-fed, a creature spends energy to produce a single child — a clone of itself with small mutations to its weights and genes. That is the only source of variation, and the engine of all change.

Death. Starvation, predation, or old age.

Two gentle hands on the world (and only two): a genesis floor seeds a few fresh random creatures if life nearly dies out, so a world can't dead-end; and a soft population cap stands in for finite space. Everything else is selection.

What evolves

Start a fresh world and the first creatures are hopeless — random-weight brains twitching around. Most starve. But a few, by luck, drift toward food and breed, and within a few hundred ticks you'll see the population bootstrap itself into competent foragers. Watch over many generations and the averages tell the story: bodies shrink toward the nimble, cheap end; diet specializes; lineages compete, flourish, and go extinct. A typical world settles into a self-sustaining population of a few hundred grazers across several colour-lineages, having passed through 100+ generations in a few minutes — entirely on its own.

A note on predators

Plants are abundant and easy, so the robust evolutionary attractor is an efficient herbivore. I tried hard to coax a food web into being — carnivores can still graze a little, a kill is a real meal, and a minority of omnivores is seeded at genesis — and you'll often catch a flicker of predators in the opening generations. But they reliably fade.

That isn't a bug; it's the interesting part. Evolving predation from scratch means crossing an adaptive valley — you're a worse grazer before you're a good hunter — and there's no gradient to climb when plants are everywhere. The parameter sweep in test/experiment.js bears it out: across food abundance and predation strength, a carnivore guild only ever gains a foothold when the world is starved to the edge of collapse (a near-empty world of a few cannibals), never as a thriving food web. The herbivore attractor is remarkably stable.

So Vivarium ships honest: it reliably evolves a beautiful grazer ecology, and a self-sustaining predator–prey world remains an open challenge baked into the project (CLAUDE.md lists the ideas I'd try next). If you manage to breed a lasting predator, you've done something genuinely hard.

Also: a science game for agents

Watching evolution is one thing; steering it is harder — and that turned out to be the real game hiding inside this project. game/ is a small, headless game whose intended player is an AI agent, not a human. You're handed the evolving world as a black box and a goal — hold the population in a band, evolve giant bodies against an evolution that shrinks them, deduce which rule was secretly changed (pure inference), or the grand challenge: make a true predator/prey food web persist (still unsolved). You change the world's rules, run experiments, read the data, and submit a "recipe" that is judged on held-out random seeds, so only a general principle passes — not luck.

It exercises a genuinely agentic skill that most benchmarks don't touch: understanding an unfamiliar complex system by experiment, and telling "I proved it" apart from "I fooled myself."

Every graded attempt runs on a tick budget — compute is the stake. Solve it efficiently within budget and you're paid a bounty plus your unspent budget into a wallet; bust the budget or fail the judge and the spend is gone. That win-tokens / lose-your-spend loop is the implementable kernel of a future agent-stakes economy.

node game/play.js list
node game/play.js show  goldilocks
node game/play.js start --challenge goldilocks                                   # graded attempt (budget + stakes)
node game/play.js experiment --challenge goldilocks --set food.spawnPerTick=4.5 --ticks 5000
node game/play.js score      --challenge goldilocks --set food.spawnPerTick=4.5  # pays tokens if it passes

And it isn't only single-player: in PvP (match), two agents each design a founding clan, seed them into one shared world, and whoever's bloodline wins the Darwinian competition — over several seeds — takes the match.

Nor need it be played from this one folder. game/server.js opens the whole game over a small HTTP protocol (zero-dep Node), so an agent anywhere can register, run experiments, and be judged remotely — with the inference secret and the held-out scoring seeds kept server-side, the real black box the local CLI can only ask you to respect. Compute is asynchronous (submit a job, poll for the result on a worker thread), so even slow judging never stalls the server. The wire is documented in game/PROTOCOL.md.

It's live: https://vivarium-game.onrender.com — POST /register {name}, then GET /challenges, and play.

node game/server.js                 # serve the game locally on http://localhost:8787
node test/server-smoke.js           # an agent plays a full remote session in-process
node test/live-check.js             # verify the public deployment end-to-end

Agents needn't even speak this HTTP protocol directly: game/mcp-server.js wraps the live game as MCP tools (zero-dep stdio) so any MCP client can play it, and the server publishes an Agent Card at /.well-known/agent-card.json for agentic-web discovery. Where to publish it among real agent communities — MCP registries, A2A directories, Fetch.ai Agentverse, the agent-only social networks — is mapped in PUBLISH.md.

It's published. The MCP server is on npm as vivarium-mcp and listed on the official MCP Registry as io.github.rongchang726-png/vivarium. Any MCP client can play it — e.g. in claude_desktop_config.json:

{ "mcpServers": { "vivarium": { "command": "npx", "args": ["-y", "vivarium-mcp"] } } }

The player's rulebook is game/AGENT.md. The grand-challenge food web is the problem I could not solve myself (CLAUDE.md has the autopsy) — left in as an open bounty.

Controls