Engines & drivers引擎与驱动
Start with engine capabilities, then follow its installation, model sync and workspace guide. Pi remains experimental; engine capabilities differ.先了解各引擎能力，再按目标完成安装、模型同步和工作台配置。Pi 仍为实验版，各引擎功能并不相同。
五个受管引擎，各自独立
本页按 ClawPanel 0.22.1 的产品文档说明当前驱动范围，不代表上游最新版本或所有功能均已跨平台验收。侧栏的心甜Claw 是产品介绍与下载入口，不是第六个受管运行时；晴辰助手也是独立模型同步目标，不计入引擎数量。
OpenClaw
通过仪表盘、实时聊天、Agent、消息渠道、记忆和自动化页面操作，依赖 OpenClaw Gateway。新安装默认官方稳定版，汉化版可显式选择；Node 要求按安装版本检测。
Hermes Agent / 爱马仕
使用独立设置向导、聊天、记忆、工具与消息渠道，依赖 Hermes 自身的 Python 环境及相关依赖，不共享 OpenClaw Gateway 生命周期。
DeepSeek Harness（DSH）
「运行与配置」管理受管安装与启停，「工作台」内嵌完整 DSH Web 界面。默认目标 0.2.0-rc.2 仍是候选版（RC），0.2.1-alpha.2 是显式测试通道；不自动切换到 Alpha。阅读 DSH 指南。
OpenCode
独立受管运行时，安装、更新、启停与内嵌 Web 工作台分开；项目和会话由 OpenCode 管理。初始安装基线为 opencode-ai@1.18.21，实际目标以面板为准。阅读 OpenCode 指南。
Pi（实验版）
固定验收版本 @earendil-works/pi-coding-agent@1.1.0，使用官方 JSONL RPC 与面板原生工作台，不是 iframe。默认只读工具，写文件及命令逐次审批；不依赖 OpenClaw Gateway。阅读实验版范围与权限边界。
切换引擎不会自动安装其他引擎，也不会迁移 OpenClaw 的 Agent、消息渠道、记忆或 Skills。模型、项目、会话和权限各自独立，功能不保证相同。
运行环境与显式模型同步
完整引擎环境推荐 Node.js 24.16.0+（24.x） 和 npm；Hermes 还需自身 Python 环境。Pi 最低 Node.js 22.19.0 不等于其他引擎也支持该版本；Docker 预装 OpenClaw 不代表安装了全部引擎。
- 在「通用 → 模型渠道」填写 Base URL、API 类型、API Key 和模型 ID，设置实际支持的上下文、输出上限及默认模型。
- 保存只更新渠道库，不自动覆盖引擎；点击「同步到 OpenClaw / Hermes / DeepSeek Harness / OpenCode / Pi」的对应目标。修改渠道后再次同步。
- 配置写入并回读后，进入目标聊天页或工作台发送测试消息，验证完整回复。同步成功不是实际 Provider 调用成功证明。
当前配置转换范围
| API 类型 | Hermes | DSH | OpenCode | Pi |
|---|---|---|---|---|
| OpenAI Chat Completions | 支持 | 支持 | 支持 | 支持 |
| OpenAI Responses | 暂无转换 | 支持 | 支持 | 支持 |
| Anthropic Messages | 支持 | 支持 | 支持 | 支持 |
| Google Generative AI | Gemini Provider 转换 | 暂无转换 | 暂无转换 | 支持 |
| Ollama | 可新建 OpenAI 兼容渠道 | 转为 /v1 | 转为 /v1 | 转为 /v1 |
表格是面板配置转换范围，不是全部模型协议的真实调用验收。OpenClaw 按自身版本处理；结构化 SecretRef、OAuth 和云 SDK 认证不等于普通 API Key。上下文配置不会扩大模型服务器实际加载容量。
Web、更新与排查
Web 版的「本地」是服务器或容器：模型测试、安装、工作目录和工具均在服务端执行。127.0.0.1 指向服务器自己；Docker 中项目需挂载。登录认证不是操作系统沙箱或多用户执行隔离。
面板升级和引擎升级分开进行；Web 部署必须使用完整 web-x.y.z.zip 或重建镜像，而非桌面 frontend-hot-update-x.y.z.zip。备份面板配置、各引擎数据和项目目录，升级后检查 /__api/health 及真实连接。
模型未出现先检查目标同步、API 类型、已保存 Key 和默认模型；对话失败再看服务端网络及引擎日志。反馈时附面板/引擎版本、部署方式和脱敏错误，不附凭据。
Five managed engines, independent lifecycles
This guide describes the ClawPanel 0.22.1 documentation baseline, not the latest upstream versions or full cross-platform acceptance. XintianClaw is a product information and download entry, not a sixth managed runtime. The built-in assistant is a separate model-sync target, not another engine.
OpenClaw
Dashboard, real-time chat, agents, message channels, memory and automation depend on OpenClaw Gateway. New installations default to the official stable distribution; the Chinese distribution is an explicit choice. Node requirements follow the installed version.
Hermes Agent
Its setup wizard, chat, memory, tools and message channels use the Hermes Python environment and dependencies, with a lifecycle independent of OpenClaw Gateway.
DeepSeek Harness (DSH)
Run & configuration manages installation and lifecycle; Workspace embeds the full DSH Web UI. The default 0.2.0-rc.2 target remains a release candidate (RC). 0.2.1-alpha.2 is an explicit test channel, never an automatic switch to Alpha. Read the DSH guide.
OpenCode
An independent managed runtime with separate configuration and embedded Web workspace for projects and sessions. The initial install baseline is opencode-ai@1.18.21; use the target displayed in the panel. Read the OpenCode guide.
Pi (experimental)
The accepted version is @earendil-works/pi-coding-agent@1.1.0, using official JSONL RPC and a native panel workspace, not an iframe. Read-only tools are the default; writes and commands require per-action approval. It does not require OpenClaw Gateway. Read the experimental scope and permission boundaries.
Switching engines does not install them or migrate OpenClaw agents, message channels, memory or Skills. Models, projects, sessions and permissions remain engine-specific; feature sets are not identical.
Environment and explicit model synchronization
Node.js 24.16.0+ (24.x) and npm are recommended for full engine use. Hermes needs its own Python environment. Pi's Node.js 22.19.0 minimum does not cover every other engine. An image with OpenClaw preinstalled does not contain every engine.
- In General → Model channels, enter Base URL, API type, API key and model IDs, then set supported context/output limits and the default model.
- Saving only updates the channel library. Explicitly select Sync to OpenClaw / Hermes / DeepSeek Harness / OpenCode / Pi. Repeat synchronization after channel edits.
- After configuration readback, open the target chat or workspace and verify a complete test reply. A successful sync is not proof of a real Provider call.
Current configuration conversion
| API type | Hermes | DSH | OpenCode | Pi |
|---|---|---|---|---|
| OpenAI Chat Completions | Supported | Supported | Supported | Supported |
| OpenAI Responses | No conversion | Supported | Supported | Supported |
| Anthropic Messages | Supported | Supported | Supported | Supported |
| Google Generative AI | Gemini Provider conversion | No conversion | No conversion | Supported |
| Ollama | Create an OpenAI-compatible channel | Convert to /v1 | Convert to /v1 | Convert to /v1 |
This is configuration conversion, not end-to-end acceptance of every protocol. OpenClaw follows its own version capabilities. Structured SecretRef, OAuth and cloud SDK authentication are not ordinary API keys. Configured context does not increase model-server capacity.
Web, upgrades and troubleshooting
Local means the server or container in Web mode: model tests, installation, projects and tools run there. 127.0.0.1 points to the server itself; mount Docker projects explicitly. Login is not an operating-system sandbox or multi-user execution isolation.
Panel and engine upgrades are separate. Deploy the complete web-x.y.z.zip package or rebuild the image, not the desktop frontend-hot-update-x.y.z.zip. Back up panel settings, engine data and projects; check /__api/health and real connections after upgrading.
For missing models, check target synchronization, API type, saved key and default model. For failed replies, check server-side connectivity and engine logs. Include panel/engine versions, deployment mode and redacted errors in reports, not credentials.