Xiaohe

Overview

Xiaohe ships as two downloadable lines: Client (desktop app) and Agent (open self-hosted package with CLI, Session API, community skills, and embedded web UI) — without giving up data ownership. Both build on industry agent harnesses and model APIs; we don’t maintain a base model or a coding-agent runtime of our own, which keeps the platform thin and easy to self-host.

Why we built Xiaohe

We are engineering researchers — computational mechanics and particle simulation. Agents became part of our daily work early. The more we used them, the more obvious a gap became: the ecosystem assumed you were already comfortable with a terminal, a model provider, and a config file. Three kinds of people were underserved, and depending on the day, we were each of them.

• Daily useSome people just want an agent: a document drafted, a deck turned into a narrated video, a PDF cleaned up — without a command line or an API key. So Xiaohe is, first of all, an agent you can talk to, in a browser or a desktop app, with install, keys, and config handled for you.
• TechnicalWe run Claude Code, Codex, and Kimi Code every day. Each is capable, and each is its own command line — its own sessions, its own config, its own way to pick a model. Xiaohe wraps a single CLI over those harnesses and turns everyday chores — reimbursement paperwork, review comments, document formatting — into skills, and into apps with voice where a command line still gets in the way. One thing existing agents don’t offer is a server of your own: the Agent is that server, self-hosted on your machine; the Client is how you reach it — the same way from any terminal, locally or across a private network.
• SimulationSimulation is our own field. Today Xiaohe works around the analysis — drafting case setup, generating inputs, turning results into reports. Next, we are moving the agent inside the analysis: a companion that helps set up, run, and interpret simulations, embedded in our particle and multiphase platform.

One honest boundary: Xiaohe is not a foundation model and not a coding-agent runtime of our own. It runs on industry agent harnesses and model APIs — our contribution is the layer that makes agents useful for real engineering work.

Components

• CLIAgent sessions, provider switching, and workspace templates in the terminal.
• Session APIREST endpoints for integrating agents into your own tools and pipelines.
• Web UIA browser chat surface for day-to-day use and shared access.
• DesktopNative shell with local services; Pro unlocks advanced client features.
• Skills HubManaged catalog of skills — authorized via your account, not a download pack.
• Model GatewayOne API key across DeepSeek, Kimi and Claude — an Anthropic-compatible /v1/messages endpoint with per-model routing and token-based billing.

Self-hosted

Run on your own machines or private cloud. Bring DeepSeek, Kimi, OpenAI, Claude, and other providers. Keep transcripts and tooling inside your network.

Pro & Skills Hub

Desktop installers are open to download. Pro client features and Skills Hub access are unlocked with credentials from the account center — activate inside the app after install.