title	xihe — General-purpose Agent Runtime Platform
category	dev-guide
sidebar_group	Overview
sidebar_order	1
lang	en

xihe — General-purpose Agent Runtime Platform

xihe (羲和) — the sun goddess in Chinese mythology who drives the sun chariot across the sky.

Give agents the status and constraints of a human.

Multi-agent orchestration · Tool invocation · Sandbox execution · Access control

⚠️ Development Status: This project is under active development. APIs and data structures may change. Not recommended for production use.

Philosophy

An agent is not a tool — it's a new role on the team.

An agent shares many similarities with a new hire: it has defined responsibilities, needs resource access permissions, can make mistakes in uncertain situations, and requires trust, supervision, and retrospection. xihe's design philosophy is built on this analogy:

Concept	People Management	xihe's Agent Management
Onboarding	Assign desk and account	Create workspace, assign Runtime instance
Responsibility	Job description	Agent prompt + role definition
Permission	RBAC permission matrix	CP access control + MCP tool-level authorization
Workspace	Physically isolated office	namespace/cgroup/seccomp sandbox isolation
Mistakes	People make mistakes, processes backstop	Agent uncertainty backed by sandbox, logs auditable
Collaboration	Team-to-team coordination	Multi-agent orchestration + human-agent collaboration

This is not a metaphor — it's the first principle of the product architecture: when agents start acting like humans, they also need to be constrained like humans. Not to limit their capabilities, but to give them a trustworthy boundary within which they can operate freely.

Architecture

Based on the above philosophy, xihe models "human" and "agent" as two equal operating entities, each expressed as an independent module, centrally orchestrated and constrained by the Control Plane (CP):

%%{init: {'theme: 'neutral'}}%%
flowchart LR
    subgraph Entities
        UI["UI<br/>(Human)"]
        AG["Agent<br/>(Agent)"]
    end

    subgraph Management
        CP["Control Plane<br/>Auth · Access · Audit · Routing"]
    end

    subgraph Infrastructure
        RT["Runtime<br/>Sandbox · Filesystem · Process"]
    end

    UI --> CP
    AG --> CP
    CP --> RT

Loading

Module	Role	Tech Stack
UI	Human interface	Vue 3 + Vite + Tailwind v4
Agent	Agent orchestration	Python + FastAPI + LangChain
Control Plane	Auth, routing, config, audit	Java + Spring Boot 4
Runtime	Sandbox execution	Rust + Axum

Quick Start

# Install dependencies (all modules)
mise run setup

# Start all services (Docker Compose backend + UI on host)
mise run dev:full

# Run full validation (lint + typecheck + build + test)
mise run validate

License

Apache 2.0