在自己的电脑上运行模型
根据设备推荐兼容模型,查看内存需求并下载安装。本地核心无需账号、订阅或 API Key。
01 / PRODUCT CAPABILITIES
根据设备推荐兼容模型,查看内存需求并下载安装。本地核心无需账号、订阅或 API Key。
在同一会话里读取、搜索和编辑文件,调用终端与浏览器工具,并查看文件改动。
把任务的不同部分交给子智能体,查看各自的执行进度,必要时单独停止。
内置 Word、PDF、演示文稿与电子表格 Skills,支持资料整理、文件生成和编辑。
保留本地会话与记忆,跟踪任务目标;为周期工作设置运行计划,查看执行历史和通知。
通过 MCP、插件和自定义模型接口扩展能力;用计划或审批模式决定 AI 可以做什么。
02 / HOW IT WORKS
模型负责理解和规划,Rust Agent Runtime 负责工具执行与权限处理,工作空间把文件、会话和结果组织起来。本地模型优先,云端模型与外部工具按需连接。
会话 · 文件 · 企业资料
工具调用 · 子任务 · 审批
文档 · 修改记录 · 执行历史
可扩展能力:Skills / MCP / 插件 / 自定义模型接口
03 / WORKFLOWS
以下为流程示意,不是客户案例或产品界面截图。
示例任务:整理项目资料中的要求与分歧,生成一份带来源的汇总草稿。
示例任务:围绕一个研究问题,分别整理资料和分析观点,再汇总成报告。
示例任务:周期性整理授权目录中的新增材料,生成待检查清单。
04 / GET STARTED
按项目文档确认运行条件,先完成一个可复现的任务。
按系统选择官方发布版本,核对安装条件。
下载受支持的模型,或配置模型接口,并指定工作目录。
连接需要的 Skills、MCP 或插件,选择适合的审批模式。
从一个真实任务开始,检查输出文件、工具操作与执行记录。
本地推理与会话存储在设备上;启用云模型、远程工具或生成服务时,完成该操作所需的数据会发送至对应服务。Zeraix 的 Rust Agent Runtime 负责工具执行;桌面通用推理目前使用 llama.cpp 系运行时,Imparo 尚非默认内置引擎。
工作空间能力:在同一工作空间处理文件与文档,通过 Skills、MCP 和插件扩展工具。 拆分任务给子智能体,查看执行过程,并按需停止单个任务。 保留会话与记忆,设置周期任务,并通过审批模式控制工具操作。
本地优先:本地核心无需 Zeraix 账号、订阅或 API Key。完成首次模型和运行时下载后,本地模型与本地工具可以离线使用;也支持按需配置自定义模型接口。
运行环境:支持 Apple Silicon macOS 13+ 与 Windows 10/11 x64,建议 16 GB 或以上内存。模型支持与可用上下文取决于设备及运行时。安装包功能以对应发布说明为准,main 分支可能包含尚未发布的更新。
产品资料核对:2026-10-21。基于官方仓库与论文;main 分支与发布安装包可能存在差异,使用前请查看对应版本说明。
下载桌面应用,选一个模型,从一个真实任务开始。
LOCAL-FIRST · OPEN SOURCE
A local-first, open-source agent workspace. Run models on your computer and let AI work with files, use tools and carry out multi-step tasks. Connect cloud models and services when you need them.

01 / PRODUCT CAPABILITIES
Get device-aware model recommendations, check memory needs and download compatible models. No account, subscription or API key is needed for the local core.
Read, search and edit files, use terminal and browser tools, and inspect file changes in the same conversation.
Delegate subtasks, follow each agent’s progress and stop agents individually when needed.
Built-in Skills support Word documents, PDFs, presentations and spreadsheets in the workspace.
Retain local conversations and memory, track goals and schedule recurring work with run history and notifications.
Connect MCP servers, plugins and model endpoints. Plan and approval modes let you control what the agent can do.
02 / HOW IT WORKS
Models interpret and plan. The Rust Agent Runtime executes tools and handles permissions. The workspace brings files, conversations and results together, with optional cloud models and external tools.
Conversations · files · knowledge
Tools · subtasks · approvals
Documents · diffs · run history
Extensions: Skills / MCP / plugins / custom model endpoints
03 / WORKFLOWS
Illustrative workflows, not customer case studies or product screenshots.
Example: summarize requirements and disagreements across project files with source references.
Example: research a topic through separate collection and analysis subtasks.
Example: periodically organize new files in an authorized directory.
04 / GET STARTED
Check the documented requirements and begin with a reproducible task.
Choose an official release for your operating system.
Prepare a supported model or endpoint and a working directory.
Connect Skills, MCP or plugins and choose an approval mode.
Start with a real task and inspect outputs, actions and run history.
Local inference and conversations stay on the device. Cloud models, remote tools and generation services receive the data needed for their actions. The Rust Agent Runtime handles tools; general desktop inference currently uses llama.cpp-based runtimes. Imparo is not yet the bundled default engine.
Workspace capabilities:Work with files and documents, and extend tools through Skills, MCP and plugins. Delegate to sub-agents, inspect execution and stop individual tasks. Keep conversations and memory, schedule recurring tasks and control tool actions through approval modes.
Local first:The local core needs no Zeraix account, subscription or API key. After initial model and runtime downloads, local models and local tools can work offline. Custom model endpoints are optional.
Requirements:Apple Silicon macOS 13+ or Windows 10/11 x64; 16 GB or more of memory is recommended. Model support and context length depend on the device and runtime. Installer features follow release notes; main may contain unreleased changes.
Product sources reviewed on 2026-10-21 against official repositories and the paper. Main and released installers can differ; check the relevant release notes.
Get the desktop app, choose a model and try a real task.