Gen-Zero: A Deterministic Decision Engine and MCP Server for AI Agents, Built in Rust

Published · AI Daily — AI-assisted deep research, methodology & disclosure

Gen-Zero is an open-source, pure-Rust cognitive decision engine and agent runtime SDK. Its reflex plugins skip token generation entirely: they map a feature vector straight to a probability distribution over named candidates, with predict latency measured at p50 164 µs and p99 238 µs (single thread, 1024-dim input, release build, this repository's own benchmark). Weights are hot-patchable, so hash-verified differential patches can swap a plugin under concurrent readers with no downtime.

For decisions that need lookahead, Gen-Zero simulates candidates in a latent world model and searches with MCTS, MPC/CEM, or A*, behind a fail-closed safety gate: a missing model, an invalid asset, or an unassessable risk produces an explicit refusal or confirmation request, not a guess. The whole runtime is one Rust binary with a CLI and a native MCP server over stdio or SSE/HTTP, so Claude Code, Codex, and other MCP clients can call it directly. It is Apache 2.0 licensed.

Why agents need a deterministic decision engine

Most agent frameworks make every choice by generating tokens: slow under load, hard to make repeatable, and expensive to audit. Gen-Zero takes a different path for choices that need to be fast and consistent. Its reflex plugins are contractive low-rank recurrent operators that map a feature vector straight to a probability distribution over named candidates, with no autoregressive generation involved. Measured on this repository's own benchmark (single thread, 1024-dim input, release build), predict latency comes in at p50 164 µs and p99 238 µs.

MCP-native, one binary

Gen-Zero ships as a single Rust binary with a CLI and a native Model Context Protocol server, over stdio for local tools or SSE/HTTP with bearer-token auth for a hosted service. Claude Code, Codex, and Google Antigravity can each add it as an MCP server in one command. The server registers `zero` and `pipeline` tools that route through the same cognitive runtime as the CLI.

Reflexes for speed, search for depth

Not every decision is a reflex. When a choice needs lookahead, Gen-Zero simulates candidate actions in a latent world model and searches with MCTS, MPC/CEM, or A*, depending on the task. Reflex plugins are also hot-patchable in production: differential patches are hash-verified before and after applying, and swap a plugin's weights under concurrent readers with zero downtime, using AVX2 acceleration where the CPU supports it.

Fail closed, not fail invisible

A safety gate sits in front of every decision. If a model is missing, an asset fails validation, or a request's risk can't be assessed, Gen-Zero returns an explicit refusal or a confirmation request instead of a plausible-looking guess. An agent that can say “I'm not sure” is safer than one that always answers.

Quickstart

git clone https://github.com/xmond/gen-zero.git
cd gen-zero
cargo build --release
./target/release/gen-zero serve --mode stdio

Or add the hosted MCP server directly:

claude mcp add gen-zero -- "https://api.gen-zero.ai/sse?token=gz_public_free"

Gen-Zero is Apache 2.0 licensed. Source, docs, and benchmarks: https://github.com/xmond/gen-zero. Official site: https://gen-zero.ai/

Sources