Monte Carlo Tree Search for reality. One scenario, a searched tree of timelines, audited.
AlphaGo searched a tree of moves to find the strongest one. WorldFork applies the same idea to social scenarios — agents expand promising decision points, simulate forward, and a god-agent backs up which paths survive.
# Run this command to install the WorldFork skill, then use its setup module to set up WorldFork: npx skills add Hilo-Hilo/WorldFork/skills/worldfork --all
Most simulators sample one trajectory. WorldFork searches the tree.
// interesting question — "what would have happened if" — is unanswerable.
// the tree is the answer — distributions, counterfactuals, audited paths.
Big Bang → Multiverse → Tick → Audit → Report.
Big Bang
Define initial conditions: actors, world rules, decision schemas, expansion constraints. WorldFork compiles this into the runnable scenario graph — the root of the search tree.
- ›actors[]
- ›world.rules
- ›decision_points[]
- ›audit.constraints[]
Designed for agents. The CLI ships today; the web UI is next.
WorldFork is operated by other agents and by humans who think like them. Stable Python CLI, stable HTTP API. Everything is reproducible with a scenario id and a seed.
A real scenario the project ships with.
The Atlas Resilience Crisis
A 40-million-person coastal megaregion enters a 180-day emergency after a heat wave, water-pressure failures, rolling blackouts, supply bottlenecks, and a climate-migration surge collide with fragile public trust and a newly deployed AI-assisted civic coordination system.
The simulation question: which governance choices preserve legitimacy, reduce harm, and prevent social fragmentation when scarcity, institutional overload, misinformation, and emergency technology all interact at once?
- ACooperative resilience success
- BOpaque emergency regime
- CMisinformation spiral
- DCivic-tech reform path
- ERegional fragmentation
- FMutual-aid federation
- GCourt-driven legitimacy reset
- HLabor & housing conflict
// same Day 0. eight plausible terminal states.
// the report compares them with full lineage.
// a stripped-down MCTS loop, on the page. expansion grows a tree, the god-agent backs scores up the tree, low-value branches drop. then it starts over.
Two endings, one root. Diff them line-by-line.
Pick two timelines. Read the world-state delta.
Same Day 0. Different decisions. The runtime persists every tick of every multiverse, so any two endings can be compared key-by-key with full lineage. Tap a chip to pick the “A” or “B” outcome.
// ↑ better in B · ↓ worse in B · · unchanged
// same root scenario · audited per tick
Paste one prompt into your agent. Done.
# Run this command to install the WorldFork skill, then use its setup module to set up WorldFork: npx skills add Hilo-Hilo/WorldFork/skills/worldfork --all
the skill routes your agent through setup, CLI, debug, reporting, and docs modules; setup walks you through prerequisites, .env config, CLI install, docker compose, migrations, seeding, readiness checks, and the onboarding demo.
- 01paste — drop the prompt above into your agent
- 02install — agent runs npx skills add ... --all
- 03follow — agent walks you through prerequisites + .env
- 04verify — agent confirms readiness and runs the demo
Common questions, plainly answered.
- Is WorldFork a SaaS product?
- No. WorldFork is open-source backend infrastructure that you self-host. There is no hosted service, no login, and no data is collected by this site.
- What is a Big Bang?
- The root scenario document that seeds a run. It defines the initial world state, the primary simulation question, and the conditions agents tick forward from. See the Atlas example.
- Does it require a specific LLM?
- No. WorldFork is model-agnostic — bring whichever model you prefer. The example scenario recommends a cheap, fast model for smoke runs and a stronger one for full demonstrations.
- What license?
- Apache License 2.0. See the LICENSE and NOTICE files in the repository for the full text.
- How do I install it with a coding agent?
- Paste the prompt at the top of this page into Claude Code, Cursor, or any coding agent. The official worldfork skill handles repo clone, environment bring-up, and first-run validation.
- Why frame this as Monte Carlo tree search?
- The four-phase MCTS loop maps cleanly onto what the runtime already does: selection picks which multiverse to advance, expansion creates children at consequential decision points, simulation ticks each child forward, and the god-agent backs up verdicts so the surviving tree reflects what actually held up under audit.
- How does expansion (branching) work?
- The runtime evaluates each tick for consequential decision points, scores them, and expands above a configurable threshold. Children inherit tick history, so each child timeline diverges from a known shared past.
- What does the audit produce?
- Persisted runtime checkpoints, every LLM call, jobs and logs, manual interventions, and structured per-multiverse + final Big Bang reports. Tied back to durable state so any run is reproducible and inspectable.
