bevy_rig

bevy_rig is a Bevy ECS integration layer for Rig.

It models providers, models, agents, tools, contexts, runs, sessions, and workflows as Bevy entities and components, then supplies systems for turning those entities into executable Rig requests.

The crate is aimed at applications that want to keep AI orchestration inside Bevy's ECS instead of building a separate runtime around ad hoc structs and service layers.

What it provides

provider and model registries as ECS data
agent, tool, and context entities
session and transcript persistence inside the world
run preparation, execution, streaming, and commit system sets
workflow graph entities and execution helpers
diagnostics helpers for inspecting runtime state

Installation

[dependencies]
bevy_app = "0.18.1"
bevy_ecs = "0.18.1"
bevy_tasks = "0.18.1"
bevy_rig = "0.1.0"

Optional features:

media: enables Rig audio, image, and PDF support
mcp: enables Rig MCP support
experimental: enables Rig experimental APIs

Quick start

use bevy_app::App;
use bevy_ecs::prelude::*;
use bevy_rig::prelude::*;

fn main() {
    let mut app = App::new();
    app.add_plugins(BevyRigPlugin);
    app.add_systems(RunExecution, complete_echo_runs.in_set(RunExecutionSystems));

    let agent = {
        let world = app.world_mut();
        spawn_agent(world, AgentSpec::new("echo-agent", "mock-echo")).agent
    };

    app.world_mut()
        .write_message(RunAgent::new(agent, "hello from bevy_rig"));
    app.update();
}

fn complete_echo_runs(
    mut commands: Commands,
    runs: Query<(Entity, &RunRequest, &RunStatus), With<Run>>,
) {
    for (run, request, status) in &runs {
        if *status != RunStatus::Queued {
            continue;
        }

        mark_run_completed(&mut commands, run, format!("Echo: {}", request.prompt));
    }
}

Core concepts

Providers and models

Providers and models are spawned as entities and indexed through registries. Agents bind to models through component references instead of direct ownership.

Agents and contexts

Agents, tools, and context documents are separate entities. This makes it easy to inspect, retarget, attach, detach, and query them through normal ECS systems.

Runs and sessions

Runs represent individual execution requests. Sessions and chat messages persist conversation state inside the Bevy world.

Workflows

Workflows are graph-shaped ECS data. Nodes, edges, and execution state are all represented with components and events, which makes them easy to debug and visualize.

Examples

The crate includes small headless examples under examples/:

headless_echo
provider_models
tool_dispatch
workflow_graph
workflow_execution
rig_provider_run
rig_provider_workflow

Run one with:

cargo run --example headless_echo

Current scope

bevy_rig focuses on ECS modeling and orchestration. It does not provide a UI layer, an opinionated agent product shell, or a full provider credential management system.

bevy_rig 0.1.0