MIT licensed and AI-native
Build worlds where intelligent agents are a core system, not an afterthought.
A free, MIT-licensed Rust game engine where AI agents are a first-class runtime system. Deterministic simulation, tool-validated behavior, and Criterion-backed performance — inspectable, reproducible, and open.
Scope
Deterministic ECS, AI planning and validation, replay-safe simulation.
12,700+ agents at 60 FPS
977 Miri tests, 0 undefined behavior
71+ Kani proof harnesses
Quick navigation
Jump straight to the area you need.
Set up the workspace and run your first system.
- Quick startOpen quick start
- InstallationInstall dependencies
- First companionBuild your first AI companion
Understand the runtime model.
- OverviewRead architecture overview
- AI-native designInspect AI-native design
- ValidationStudy tool validation
Explore the engine subsystems.
- AI systemOpen AI systems
- Physics and fluidsInspect simulation systems
- Terrain and navigationBrowse world systems
Demos, benchmarks, and optimization.
- ExamplesBrowse examples
- BenchmarksOpen benchmark dashboard
- OptimizationRead optimization guide
Contribute and build from source.
- ContributingOpen contribution guide
- BuildingBuild from source
- TestingReview validation workflow
Subsystem APIs and crate documentation.
- API overviewOpen API reference
- Crate mapBrowse crate documentation
- CLI and configInspect tools and configuration
Why AstraWeave
Architecture, safety, and performance you can verify.
Perception → reasoning → planning → action, built into the runtime.
World snapshots, plan intents, tool validation, behavior trees, GOAP, and LLM-backed planning all plug into a deterministic loop.
Bit-identical replay, validator-gated actions, and formal verification.
Core unsafe paths are exercised under Miri. Kani proofs back critical ECS, math, and SDK invariants. Replay validation is built in, not bolted on.
Every performance claim links to a reproducible Criterion measurement.
ECS, AI planning, physics, rendering frame times, and SIMD throughput are all measured with specific numbers, not broad adjectives.
Engine loop
Perception → reasoning → planning → validation → action.
Perception
World snapshots from deterministic ECS state.
Reasoning
Behavior trees, utility systems, GOAP, or LLM logic.
Planning
Action sequences with costs, priorities, and fallbacks.
Validation
Cooldowns, LOS, pathing, and sandbox constraints.
Action
Approved commands flow into simulation, physics, and rendering.
What ships today
A focused stack for intelligent, simulation-heavy games.
Six validated modes plus hybrid arbiters.
Classical planners, behavior trees, utility logic, LLM orchestration, ensemble patterns, and hybrid arbiters.
Ordered simulation for replay, tooling, and scale.
Archetype storage, system staging, iteration guarantees, and event channels form a reproducible simulation backbone.
wgpu-based rendering with real engine workloads.
PBR materials, clustered lighting, GPU skinning, post-processing, and LOD tooling.
Character control, spatial hashing, fluids, and more.
Collision, character motion, destructibles, ragdolls, vehicles, gravity zones, and fluid simulation.
Navmesh, terrain, scene streaming, and gameplay layers.
Terrain generation, navigation meshes, crafting, quests, dialogue, and procedural content.
Example suite, editor tooling, and a C ABI for embedding.
A large example suite, editor workflows, and a stable C SDK layer.
Current measurements from the codebase.
- Agent capacity at 60 FPS12,700+
- AI validation throughput6.48M checks/sec
- Frame time at 1,000 entities2.70 ms
- ECS world creation25.8 ns
- Character move cost114 ns
- SIMD batch over 10k entities9.879 us
Top Benchmarks · Latest Run
Criterion.rs statistical benchmarking · Feb 2026
Bake 10K Triangles
E2E Plan Gen · Cache Miss
Cache Latency · 200ms
Light Binning · High
Bloom Upsample · Mip 0
Light Binning · 5K
Cache Latency · 100ms
Game Loop · 5K Stress
Verification and testing across the stack.
- Weighted line coverage59.3% across measured crates
- High-coverage crates14 crates at 85%+
- Miri validation977 tests, 0 UB
- Kani verification71+ harnesses
- Prompt mutation testing100% adjusted kill rate
- Desktop targetsWindows, Linux, macOS
Pick an entry point.
- System designOpen architecture overview
- Performance dataOpen benchmarks
- Build and runRead quick start
- Examples and demosBrowse examples
- Contributing workflowOpen contribution guide
- Reference implementationInspect Veilweaver
Use cases
Where AstraWeave fits best.
Companions, directors, and systemic encounters.
Projects where NPCs need to observe, plan, and react with more depth than state-machine scripting allows.
Validation and replayability matter.
Deterministic simulation and tool-gated actions for anti-cheat, reproducibility, and replay validation.
Benchmarkable AI-native architecture.
Test agent scale, planning strategies, and hybrid AI control under measurable conditions.
Rust core with a C ABI.
Adopt focused subsystems through the modular crate structure instead of committing to the whole stack.
Design lineages
Games this engine could help realize.
In the lineage of Dwarf Fortress or RimWorld.
Agent autonomy, world-state memory, logistics, and emergent story generation.
In the lineage of Civilization or Crusader Kings.
Multi-agent diplomacy, advisor systems, strategic planners, and explainable AI.
In the lineage of X-COM or Battle Brothers.
Tool validation, cover awareness, action planning, and replay-safe combat loops.
In the lineage of Kenshi or Mount and Blade.
Large numbers of autonomous actors with persistent world consequences.
In the lineage of Deus Ex or Dragon Age.
Companion decisions, quest reactivity, systemic encounters, and director-style orchestration.
Licensed ports or original successors.
Rebuild ambitious systemic designs with modern AI-native architecture.
Next step
Clone, build, and decide from evidence.
AstraWeave is free and MIT licensed. Evaluate, adopt subsystems, or contribute back without platform lock-in.