Inference and execution gateway for A.L.I.C.E. agents
alice-runtime: a single Bun binary that owns the agent loop, scheduler, sandboxed tool execution, and the trace→reflection→promotion learning loop. Replaces OpenClaw.
The execution engine
The runtime is what makes agents actually run. It sits between your AI models and your agents — loading their workspace files, executing tools in sandboxed environments, managing the agent loop, and tracking every decision through the reflection loop that lets agents improve over time.
alice-runtime is a single Bun binary listening on port 8788. It exposes OpenAI- and Anthropic-shaped HTTP APIs so it slots into existing infrastructure without architectural changes. It owns the full lifecycle: model calls, tool execution, session state, tracing, and the trace→reflection→promotion pipeline that promotes successful behaviors up the capability ladder.
It integrates with A.L.I.C.E. | Agents for persona definitions, RecordorAI for memory, and is dispatched by A.L.I.C.E. | Hub via the OpenAI-compatible /v1/chat/completions endpoint.
5 messaging channels, one runtime
3 inference tiers
Direct
OpenAI and Anthropic API adapters — lowest latency, full model access
LiteLLM Sidecar
Unified interface across 100+ LLMs, protocol translation, and fallback routing
Local / llama.cpp
Fully offline inference for private deployments — no external API calls
OpenAI-compatible endpoint
POST /v1/chat/completions
Authorization: Bearer <runtime-api-key>
Content-Type: application/json
{
"model": "gpt-4o",
"messages": [
{ "role": "user", "content": "Check status of node-7" }
],
"tools": [{ "type": "function", "function": ... }],
"stream": false
}What the runtime owns
Ready to run?
Runtime is part of the full A.L.I.C.E. stack. See how it fits with Hub and Agents.