Operit

AAswordmanLGPL-3.0活跃维护

Android 上能力最强的 AI Agent 与聊天软件,支持本地 llama/MNN 模型和终端操作,让手机直接跑智能体任务。

The most powerful AI agent and AI chat software on Android/Operit是一款Android上能力最为强大、发展最久的AI Agent

v1.12.2发布于 2026-09-18 04:20 UTC9,494star +65 / 7 天827fork同步于 2026-11-10 11:30 UTC

怎么上手

由 README 提炼,执行前请确认
从 Releases 页面下载最新 APK → 安装并启动 → 按引导配置模型与权限
硬件要求按 README 说明

仅支持 ARM64(arm64-v8a)设备,系统需 Android 8.0(API 26)及以上;内存与存储占用取决于终端环境、工具包和本地模型,需按各模型文档预留空间;README 未给出具体显存或内存数值。

下载:源码 zip·GitHub Releases(安装包)本站不镜像安装包

项目说明

由 README 自动翻译提炼

Android 上的开源 AI Agent,把云/本地模型接到系统能力上执行真实任务

Android 上能力最强的 AI Agent 与聊天软件,支持本地 llama/MNN 模型和终端操作,让手机直接跑智能体任务。

核心能力

  • 通过工具调用、工作流与 ToolPkg/MCP/Skills 扩展执行任务
  • 支持云端模型、自定义兼容端点与密钥池,也可本地推理
  • 内置 MNN 与 llama.cpp 运行 GGUF 本地模型
  • 授权后调用 Android 系统能力并操作应用界面
  • 手机内提供 Ubuntu 24.04 用户空间、SSH 与 SFTP

容易踩的坑

只支持 ARM64 设备,非 arm64-v8a 机型无法安装;本地推理需自行准备 MNN 或 GGUF 模型文件,占用空间按模型文档预留;Web Chat/HTTP API 默认关闭,启用前需确认网络暴露与 bearer token 设置。

内容由模型根据 README 生成,可能不完整或不准确,以官方文档为准。看原文 README

最近发版

5 个版本v1.12.2
  1. v1.12.2Operit AI V1.12.2最新正式版2026-09-18 · 2026-09-18 04:20 UTC

    本次更新重点围绕模型兼容性、工具调用、聊天稳定性和数据管理展开,同时补充 ToolPkg、统计、恢复与性能监控能力。

    增加

    • 新增 Token 用量账本、统计看板、活动视图、价格覆盖、缓存命中率、生成速度和安全删除能力
    • 新增独立 TTS/STT 服务档案,支持创建、复制、切换、删除和旧配置迁移
    • 新增按窗口或指定日期范围重建聊天记忆
    • Compose DSL 文件选择器新增文档、图片、视频、媒体、目录和相机模式
    • 新增 AiChat、AdaptiveSidePanel、Compose DSL 对话框和聊天消息长按菜单
    • 新增 DeepSeek Harness 运行时侧栏和完整 ToolPkg 运行时支持
    • 新增 MiniMax、OpenCode Zen/Go、xAI Grok 和 OpenAI Codex OAuth 提供商,并支持 Codex 配额查询和 PDF 输入
    • 新增按提供商和模型动态配置思考深度,补充 DeepSeek Responses、Qwen、Gemini 等端点规则
    • 新增 DeepSeek Responses Web Search、服务端工具记录和 reasoning 回放
    • ToolPkg 新增版本化 API、包依赖、加载顺序、运行态 Hook、消息持久化 Hook 和聊天消息菜单
    • 新增数据库、DataStore 配置健康检查和手动修复流程,并在修复前归档原始文件
    • 新增主进程、插件 QuickJS、终端和设备 CPU/内存/网络性能监控
    • 新增日语本地化、xAI 绘图自定义 API 地址、ToolPkg Logo 和已选聊天记录导出

    修复

    • 修复并行或分段工具结果在流式展示、历史记录和下一轮请求中的顺序不一致问题
    • 修复 Claude、DeepSeek、Gemini、Kimi、OpenAI 等 Provider 的工具调用与结果配对问题
    • 修复未匹配工具结果被误报为“用户取消”,并补充缺失结果的诊断信息
    • 修复 DeepSeek Responses commentary/reasoning 回放、工具历史和正文流式输出问题
    • 修复 OpenAI Responses 中 video_url/input_video 内容丢失问题
    • 修复长工具结果、超大聊天导出、被中断的流式回复和多工具结果被提前丢弃的问题
    • 修复群聊编排重复发送用户消息,以及消息变体事件使用旧快照的问题
    • 修复 Compose DSL 文本框光标跳动、复制预览回弹、输入法遮挡和 Markdown 表格链接点击问题
    • 修复 MCP 工具发现成功后未及时注册,以及并发权限描述注册的数据竞争
    • 修复媒体能力测试把“已连通”误判为“能力已验证”的问题
    • 修复浏览器桌面 viewport、截图尺寸、Provider 历史图片和模型列表刷新问题
    • 修复悬浮窗非法尺寸、初始化竞态和连续崩溃保护问题
    • 修复 Web Chat JavaScript 模块 MIME 类型错误
    • 修复备份恢复、SQLite 参数数量限制、偏好迁移和 ToolPkg manifest 解析问题
    • 修复 OpenCode、Codex、DeepSeek 和 OpenAI 的思考参数映射与历史兼容问题

    优化

    • ToolPkg 发布、API 版本展示、依赖检查、包排序、manifest 选择和隐藏目录排除更加可靠
    • 摘要支持分区配置、独立标题、启用状态和对话回顾设置
    • Token 统计的迁移、恢复、重试、价格计算和长时间运行稳定性得到提升
    • 状态卡片改用内置完整 Material Symbols 字体,离线环境下也能正常显示
    • APK 逆向工具升级至 1.1.0,JADX 支持隔离进程、串行任务和更完整的错误诊断
    • 改进聊天复制、长历史浏览、流式滚动、LaTeX 渲染和大消息处理
    • CI 拆分 Android/JVM 检查并优化缓存、原生依赖和 Nightly 构建流程
    • 市场贡献者内容、Logo、包来源、版本信息和审核流程更加完整
    • README 默认切换为英文,并保留 README.zh-CN.md 中文版本
    • 补充日语及多种语言的模型、市场、ToolPkg、统计和设置文案

    特别致谢

    感谢所有参与本版本开发、审查、测试、文档和本地化工作的贡献者。特别感谢 @AAswordman、@luojiaping、@CATMIAOZHI、@3316891527、@yoruuuchan、@YunXi-Aurora、@tuxKOH、@AikingChoice、@AnxForever、@nailuoGG 和 @okker24。

    This update focuses on model compatibility, tool execution, chat stability, and data management, while also improving ToolPkg, statistics, recovery, and performance monitoring.

    Added

    • Added token usage ledgers, dashboards, activity views, pricing overrides, cache hit rates, generation speed, and safe deletion
    • Added independent TTS/STT service profiles with migration from the previous configuration
    • Added chat-memory rebuilding by window or selected date range
    • Added document, image, video, media, directory, and camera modes to the Compose DSL file picker
    • Added AiChat, AdaptiveSidePanel, Compose DSL dialogs, and chat-message long-press menus
    • Added the DeepSeek Harness runtime sidebar and complete ToolPkg runtime support
    • Added MiniMax, OpenCode Zen/Go, xAI Grok, and OpenAI Codex OAuth providers, including Codex quota and PDF input
    • Added provider-aware and per-model thinking controls for DeepSeek Responses, Qwen, Gemini, and related endpoints
    • Added DeepSeek Responses Web Search, server-side tool records, and reasoning replay
    • Added versioned ToolPkg APIs, package dependencies, load ordering, runtime hooks, message-persisted hooks, and chat menus
    • Added database and DataStore health checks with manual repair and original-file archiving
    • Added CPU, memory, and network monitoring for the app, QuickJS plugins, terminals, and the device
    • Added Japanese localization, configurable xAI drawing base URLs, ToolPkg logos, and selected chat exports

    Fixed

    • Fixed parallel and multi-part tool results appearing in different orders across streaming output, history, and follow-up requests
    • Fixed tool-call/result pairing across Claude, DeepSeek, Gemini, Kimi, OpenAI, and proxy tool histories
    • Fixed unmatched tool results being reported as user cancellation and added clearer missing-result diagnostics
    • Fixed DeepSeek Responses commentary/reasoning replay, tool history, and normal-text streaming
    • Fixed video_url/input_video content being dropped during OpenAI Resp
  2. v1.12.1Operit AI V1.12.12026-08-08 · 2026-08-08 20:33 UTC
    查看更新说明

    本次更新重点围绕角色与记忆、插件开发、远程协议、市场发布和聊天稳定性展开。用户资料升级为 Markdown 档案,角色主题可以独立管理;ToolPkg 新增 WASM 与消息 Hook,远程 MCP 和 A2A 接入;同时修复长消息、流式渲染、取消生成、待发队列和市场修订等问题。

    增加

    • 用户资料升级为可自由编辑和预览的 Markdown 档案,并支持自动更新开关、锁定和自定义记忆提取规则
    • 角色卡和角色组可以分别保存主题;主题编辑支持草稿、目标切换确认和未保存修改保护
    • 新增独立的聊天标题生成模型配置,可根据第一条消息自动生成标题
    • 工具提示、ToolPkg 和 Skill 支持拖拽排序;长粘贴文本可以自动转换为文本附件
    • 消息复制新增“纯文本 / Markdown 源码”模式,表格、代码块、链接和公式可以转换为可读文本
    • 远程 MCP 支持 Streamable HTTP、SSE、Bearer Token、自定义 Header、工具分页和结构化结果
    • 外部 HTTP 服务新增 A2A 1.0,支持消息发送、流式输出、任务查询、取消和订阅
    • ToolPkg 新增 wasm_modules 和 ToolPkg.wasm.call(),支持 i32、i64、f32、f64 参数及多返回值,并支持 AssemblyScript abort
    • ToolPkg 新增 message_persisted 消息持久化 Hook;聊天输入、Prompt 和 Summary 前置 Hook 增加整轮时间限制,默认 10 秒,可配置为 1–60 秒
    • 新增角色卡管理工具,以及截图、聊天消息范围读取、指定终端会话操作和带地址的定位接口
    • 市场发布支持直接上传文件或选择已有 GitHub Release 文件,并支持审核退回后的修订发布
    • Worldbook 示例新增分组管理、条目复制、角色卡多绑定和分组删除处理
    • 新增罗马尼亚语,并补充多种既有语言中的市场、工作流和设置文本

    修复

    • 修复多个模型服务中思考内容丢失、工具调用 ID 不稳定、历史工具调用无法正确重放等问题
    • 修复五档思考深度在 Claude、DeepSeek、Gemini、NVIDIA、OpenRouter、Qwen 等服务中的映射问题
    • 修复切换角色卡后,发送、重新生成或记忆处理仍可能使用错误角色的问题
    • 修复取消生成时部分回复、Token 和耗时信息丢失,以及旧任务干扰新一轮生成的问题
    • 修复切换聊天或页面重组后待发送消息丢失、错发到其他聊天的问题
    • 修复超大消息和消息分支无法正常读取、导出或参与记忆处理的问题
    • 修复流式工具调用被截断、末尾内容不刷新、Markdown 节点错误复用和表格滚动位置重置的问题
    • 修复自动续写误用输入框草稿、附件和回复目标的问题
    • 修复消息图片、LaTeX 公式、小屏弹窗和暗色主题市场文字的显示问题
    • 修复 API Key 轮换、模型首次配置、语言重启恢复和损坏头像配置导致的异常
    • 修复插件导入后列表不立即刷新、市场评论权限不准确和审核状态操作混乱的问题
    • 修复终端退出后旧会话被再次复用、网页访问超时处理不明确等问题

    优化

    • 提升超长对话下的 Token 统计、上下文判断和摘要触发稳定性
    • 文件内容搜索改为应用内置实现,不再依赖终端环境预装 rg
    • 优化流式消息渲染、自动滚动、长历史浏览和中断消息保存
    • ToolPkg 发布支持 JavaScript AST 压缩、变量改名和可达模块裁剪,并增加市场来源与作者信息
    • ToolPkg 导入增加远程阻止名单检查,QuickJS 的超时、中断、容器隔离和销毁流程更加稳定
    • 远程 MCP 的工具发现、连接复用、超时处理和损坏文本拦截更加可靠
    • GitHub 登录改用 Operit OAuth broker,并支持内嵌页面和系统浏览器两种方式
    • 市场完善版本兼容提示、作者与来源展示、审核详情和发布权限控制
    • 工作流会提前准备所需工具和插件环境,并显示更明确的失败阶段与原因
    • 原生依赖、STT 模型和外部原生库增加固定版本与完整性检查
    • Release 和 Nightly APK 支持 Android V3 签名轮换,Debug 使用独立包名和签名
    • CI 改为按修改范围选择检查任务,并更新 WebChat、ToolPkg、文档和协作流程
    • 全面更新中英文 README 与产品展示图片

    特别致谢

    特别感谢 @luojiaping 对本版本的持续投入。本次更新中,多项模型兼容、聊天与长消息稳定性、用户资料、构建流程和多语言改进都离不开他的贡献。

    Added

    • Upgraded user profiles to freely editable Markdown documents with preview, automatic-update controls, locking, and custom memory-extraction rules
    • Added separate themes for character cards and character groups, with draft editing and unsaved-change protection
    • Added a dedicated title-generation model configuration for automatically naming new chats
    • Added drag-and-drop ordering for tool prompts, ToolPkg packages, and Skills, plus automatic conversion of long pasted text into attachments
    • Added Plain Text and Markdown Source copy modes with readable conversion for tables, code blocks, links, and formulas
    • Added remote MCP support for Streamable HTTP, SSE, bearer tokens, custom headers, paginated tool discovery, and structured results
    • Added A2A 1.0 support to the external HTTP service, including message sending, streaming, task lookup, cancellation, and subscription
    • Added ToolPkg wasm_modules and ToolPkg.wasm.call() with i32, i64, f32, f64, multiple returns, and AssemblyScript abort support
    • Added the ToolPkg message_persisted hook; chat-input, Prompt, and Summary pre-send hooks now have a whole-chain timeout of 10 seconds by default, configurable from 1 to 60 seconds
    • Added character-card management tools, screenshot capture, ranged chat-message reading, terminal session targeting, and optional address lookup for location
    • Added marketplace publishing through direct upload or existing GitHub Release files, together with revision submission after requested review changes
    • Added Worldbook grouping, entry copying, multi-character binding, and group-deletion handling
    • Added Romanian localization and expanded marketplace, workflow, and settings translations

    Fixed

    • Fixed lost reasoning content, unstable tool-call IDs, and broken historical tool-call replay across multiple model providers
    • Fixed five-level thinking mappings for Claude, DeepSeek, Gemini, NVIDIA, Op
  3. v1.12.0Operit AI V1.12.02026-07-01 · 2026-07-01 14:47 UTC
    查看更新说明

    本次更新重点围绕新市场、工具能力、聊天体验和稳定性展开。统一市场继续重构,新增音乐、蓝牙、系统操作和语音相关工具,聊天附件、滚动导航和历史加载更稳定,同时补充多语言、崩溃恢复、安全配置和模型兼容性修复。

    增加

    • 新增统一市场的浏览、详情、作者、管理和发布流程,市场项目、安装状态、统计、审核状态和 GitHub 发布链路更完整
    • 新增音乐播放工具和播放队列工具,支持通过工具调用控制音乐播放,并同步补充 JS/TypeScript 类型声明
    • 新增蓝牙设备工具和更多系统操作工具,扩展 system_tools 示例包和工具开发文档
    • 新增 Mimo 语音、Doubao TTS、MIMO TTS 声音克隆和 TTS 排队播放能力
    • 新增工具 Hook 拦截能力,工具包可以参与工具生命周期处理
    • 新增长 MCP 结果保存为文件的能力,避免超长工具结果直接挤占聊天内容
    • 新增崩溃后的数据恢复流程和启动隔离机制,异常后更容易进入修复界面
    • 新增自定义总结提示词设置,并支持在设置页中自动保存
    • 新增 HTTP/WebUI 自定义 bearer token 设置,方便外部访问时配置鉴权
    • 新增 Claude 1 小时 prompt cache 配置、群组编排模式和基于时间戳的聊天历史加载
    • 新增韩语本地化,并继续补充西班牙语、葡萄牙语等语言资源

    修复

    • 修复 grep_code 在复杂目录中过度处理 ripgrep 结果的问题
    • 修复 Claude Opus 4.8/4.7 thinking 参数不兼容导致请求失败的问题
    • 修复云端 embedding endpoint 校验不准确的问题
    • 修复 DashScope 兼容模式音频识别可能返回 400 的问题
    • 修复自定义总结提示词不会自动保存的问题
    • 修复默认配置中硬编码 DeepSeek API Key 的安全问题
    • 修复进程限制页面按钮文字可能溢出的问题
    • 修复模型提供商显示名称不一致的问题
    • 修复中文文案中重复出现“未”的问题
    • 修复新市场发布、统计、审核状态、详情页和多语言文案中的多处问题

    优化

    • 聊天滚动到底部功能整合进滚动导航器,Android 和 Web Chat 的长历史浏览体验更一致
    • 附件状态管理重构,重复附件文件名会被去重,图片附件处理也更稳定
    • 默认显示角色名和用户名,多角色聊天、群组编排和历史定位更清晰
    • 主题设置页拆分为多个 Tab,背景、基础、聊天、输入和界面设置更容易维护
    • 工具选择弹窗增加工具描述,管理员更容易理解每个工具的用途
    • Markdown、表格、代码块和文本选择体验继续改进,长内容展示更稳定
    • 工具包运行时管理拆分为加载器和管理器,工作区变更追踪与附件处理更清晰
    • 模型服务、OpenAI/Qwen/Claude 调用、结构化工具调用和消息编辑流程继续调整
    • 构建文档补充 JDK 21、pnpm workspace 和 clone 构建类型等说明

    This update focuses on the new marketplace, tool capabilities, chat experience, and stability. The unified marketplace has been further rebuilt, new music, Bluetooth, system-operation, and voice tools have been added, chat attachments, scroll navigation, and history loading are more stable, and this release also includes localization, crash recovery, security, and model-compatibility fixes.

    Added

    • Added unified marketplace flows for browsing, details, author pages, management, and publishing, with more complete project, installation-state, statistics, review-state, and GitHub publishing support.
    • Added music playback and play-queue tools, allowing playback control through tool calls, with matching JS/TypeScript type updates.
    • Added Bluetooth device tools and more system-operation tools, expanding the system_tools example package and tool-development documentation.
    • Added Mimo voice, Doubao TTS, MIMO TTS voice cloning, and queued TTS playback.
    • Added tool Hook interception so tool packages can participate in the tool lifecycle.
    • Added saving long MCP results to files, preventing oversized tool results from directly crowding chat content.
    • Added crash-repair data recovery flow and startup isolation, making it easier to enter a repair screen after failures.
    • Added custom context-summary prompt settings with automatic saving in the settings page.
    • Added custom bearer token support for HTTP/WebUI access.
    • Added Claude 1-hour prompt cache configuration, group orchestration mode, and timestamp-based chat-history loading.
    • Added Korean localization and continued improving Spanish, Portuguese, and other language resources.

    Fixed

    • Fixed grep_code over-processing ripgrep results in complex directories.
    • Fixed Claude Opus 4.8/4.7 thinking-parameter incompatibility that could cause request failures.
    • Fixed inaccurate validation for cloud embedding endpoints.
    • Fixed DashScope compatibility-mode audio recognition potentially returning 400 errors.
    • Fixed custom context-summary prompts not being saved automatically.
    • Fixed the security issue of a hardcoded default DeepSeek API key in default configuration.
    • Fixed button text overflow on the process-limit remover screen.
    • Fixed inconsistent model-provider display names.
    • Fixed a duplicated Chinese word in UI text.
    • Fixed multiple issues in the new marketplace around publishing, statistics, review status, detail pages, and localized text.

    Optimized

    • Integrated scroll-to-bottom into the chat scroll navigator, making long-history browsing more consistent between Android and Web Chat.
    • Refactored attachment state management, deduplicated r
  4. v1.11.0Operit AI V1.11.02026-05-16 · 2026-05-16 11:45 UTC
    查看更新说明

    本次更新重点围绕插件生态展开,增加插件市场、更多的接口,同时增加qqbot对接等插件示范,并增加软件的稳定性。

    增加

    • 新增计划模式,计划准备阶段不会修改文件,在确定需求以及用户批准后,进行计划实现
    • 升级世界书,支持国际化、变量系统、导入、搜索、刷新、角色卡绑定、注入位置设置和更完整的管理界面
    • 新增 QQ Bot,可以直接在qq中发消息并接收到Operit的执行结果
    • 增强外部调用能力,现在可以在同一局域网里面,通过其他设备的浏览器访问Operit聊天界面了
    • 聊天输入增加@以及/来引用文件、工具包的能力
    • 聊天历史支持消息定位预览、按时间戳跳转、消息收藏以及更稳定的长历史浏览
    • 作品市场升级为项目/版本节点结构,支持输入需求,一键创建自己的插件,支持基于插件市场的插件二次修改,并重新发布作为该插件的新版本。
    • ToolPkg 支持工作流模板、工作区模板、桌面小组件、WebView UI、聊天输入 Hook、聊天视图 Hook、Prompt Hook、XML 渲染插件和跨运行时 IPC
    • Compose DSL 大幅增强,支持 slots、图片缩放、文件选择、WebView 控制器、重新渲染、文本输入同步和更多 Material 图标
    • 新增 4Router、Mimo 等模型/服务入口,并支持 ToolPkg 注册 AI Provider,增加对deepseek v4的思考适配
    • 文件工具新增更清晰的 create_file / edit_file,zip 工具支持严格的 include_root_directory 参数
    • GitHub 登录支持外部 OAuth 流程
    • 新增多个侧边栏和网页示例,包括账户本、模型站点、OpenCode、SillyTavern、Dino Runner 等,已发布于插件市场
    • 聊天消息现在可以显示更完整的 token、耗时和消息时间信息,方便排查模型调用表现

    修复

    • 修复 various_search 视频搜索可能返回首页或无效链接的问题,现在会尽量返回真实目标链接
    • 修复大聊天历史导入/导出容易占用过多内存的问题,聊天备份改用流式处理
    • 修复长聊天记录下消息定位、滚动导航、插入总结和变体加载不稳定的问题
    • 修复 Compose DSL 文本框聚焦、外部值更新和重新渲染时可能不同步的问题
    • 修复市场页面返回后已安装状态、列表数据和 MCP 插件状态不刷新的问题
    • 修复工具结果过长时直接挤占上下文的问题,最终工具结果会更克制地截断
    • 修复 zip 与调试器文件工具中的路径引用和根目录打包行为问题
    • 修复未绑定工作区统计不准确的问题,内部/外部工作区会分开识别
    • 修复浏览器工具在 iframe、点击、会话上下文和动作 settle 场景中的若干不稳定问题
    • 修复 TTS 播放,增加朗读的暂停/继续能力
    • 修复工作区的刷新逻辑,现在预览html只会在文件有改动的时候刷新

    优化

    • 删除任务完成和等待标记,记忆总结改为按照数量和时间累计总结,删除错误的"记忆附着"行为说明
    • 市场浏览、详情、管理、安装和发布流程更统一,优化市场体验
    • 绘图类插件迁移到各自的插件存储目录,减少不同插件之间的文件互相影响
    • 新增OpenAI Local,优化Llamacpp、mnn的本地工具调用方式
    • 部分内置的包已删除并发布到市场,有需要可以自行下载
    • 优化软件侧边栏,优化软件LOGO

    This update focuses mainly on the plugin ecosystem. It adds the plugin marketplace, expands plugin interfaces, includes examples such as QQ Bot integration, and improves overall app stability.

    Added

    • Added Plan Mode. Files are no longer modified during planning; implementation starts only after the requirements are confirmed and approved by the user.
    • Upgraded Worldbook with internationalization, a variable system, import, search, refresh, character-card binding, injection-position settings, and a more complete management interface.
    • Added QQ Bot, allowing messages to be sent directly in QQ and Operit's execution results to be received there.
    • Improved external access. Operit's chat interface can now be opened from a browser on another device in the same local network.
    • Chat input now supports using @ and / to reference files and tool packages.
    • Chat history now supports message-location previews, timestamp jumps, message favorites, and more stable browsing for long histories.
    • Upgraded the Artifact Market to a project/version-node structure. Users can enter requirements to create their own plugins with one click, modify plugins from the marketplace, and republish them as new versions of those plugins.
    • ToolPkg now supports workflow templates, workspace templates, desktop widgets, WebView UI, chat-input hooks, chat-view hooks, Prompt hooks, XML rendering plugins, and cross-runtime IPC.
    • Compose DSL has been greatly enhanced with support for slots, image scaling, file selection, WebView controllers, rerendering, text-input synchronization, and more Material icons.
    • Added model/service entries such as 4Router and Mimo. ToolPkg can now register AI Providers, and thinking adaptation for DeepSeek V4 has been added.
    • File tools now include clearer create_file / edit_file tools, and the zip tool now supports a strict include_root_directory parameter.
    • GitHub login now supports an external OAuth flow.
    • Added multiple sidebar and web examples, including Account Book, Model Sites, OpenCode, SillyTavern, and Dino Runner. These have been published to the plugin marketplace.
    • Chat messages can now show more complete token, duration, and message-time information, making it easier to inspect model-call behavior.

    Fixed

    • Fixed an issue where various_search video search could return homepages or invalid links. It now tries to return the actual target links.
    • Fixed excessive memory usage when importing or
  5. v1.10.1Operit AI V1.10.12026-04-17 · 2026-04-17 09:44 UTC
    查看更新说明

    新增更完整的内置浏览器与浏览器脚本支持、主题液态玻璃效果与侧边栏自定义、FBX 虚拟形象与升级后的 MMD 预览、配置编辑器调试和自动编写 Operit 插件、工作区重命名、历史跳转与双向分页,以及更多模型、语音、绘图和工作流能力。

    增加

    • 支持 FBX 虚拟形象,并升级了 MMD 预览体验
    • 支持通过本地 HTTP 接口从外部发起对话,提供健康检查、流式回调,并可按需拉起悬浮聊天
    • 模型配置现在支持单独设置自定义请求头,接入特殊网关或第三方服务更方便
    • 角色卡可以限制可用工具,角色卡和角色组也可以拥有各自独立的自定义表情
    • 系统分享文件或链接时,可以直接选择发到当前会话、最近会话或新会话
    • 内置浏览器能力大幅增强,支持标签页、历史、书签、权限申请、多窗口、最小化和视口控制
    • 浏览器脚本支持导入、安装、启停、存储和页面菜单,网页自动化和增强能力更完整
    • 支持利用配置编辑器调试和自动编写 Operit 插件
    • 新增 PDF 逐页识图、App 使用时长读取,以及应用冷启动时自动触发工作流
    • 发消息前可自动补充更多上下文,包括当前前台应用、当前页面 Activity、近期 App 使用时长、屏幕 OCR 文本等
    • 现在会自动读取工作区根目录规则文件,让项目专属约束更容易生效
    • 支持工作区重命名,并会同步更新关联聊天标题
    • 聊天历史支持快速滚动、消息跳转定位和双向分页浏览
    • 支持 GIF、HTML 换行和更稳的公式渲染,复杂消息显示效果更自然
    • 主题新增液态玻璃效果,聊天气泡、输入栏和侧边栏都支持更多外观自定义
    • 侧边栏支持独立背景色、强调色、液态玻璃按钮和展开动画设置
    • 新增更多模型和服务接入,包括 Novita、OpenRouter、MiniMax 语音/绘图,以及 Kimi Coding 端点
    • Waifu 模式支持自定义附加提示词,也可以自定义回复提示音和震动
    • 新增更多插件示例,方便扩展上下文裁剪、聊天调试等高级玩法

    修复

    • 修复工作区 @ 文件补全卡顿、候选太乱的问题,搜索更快也更准
    • 修复工具权限判断不准确的问题,package_proxy 权限检查和角色卡工具限制更可靠
    • 修复语音测试报错信息不清楚的问题,网络错误、超时和 HTTP 失败更容易排查
    • 修复 HTTP TTS 对 JSON 返回、多跳请求和 base64 音频结果的兼容问题
    • 修复 SSH 长输出和 tmux 失败导致的崩溃,超长结果现在会更稳地保存到文件
    • 修复聊天历史中的滚动定位、上下文跳转、工作流空参数、图片解析和图片上传卡住等问题
    • 修复 GIF 和文件导入时的扩展名识别问题,以及公式遮挡、重叠和 Markdown 换行不一致的问题
    • 修复 token 估算污染缓存状态、统计不准等问题,相关使用统计和缓存命中记录更完整
    • 修复 MCP 和远程服务配置崩溃、脏配置残留、插件冲突来源难定位等问题
    • 修复消息时间重复时可能导致的排序和跳转异常
    • 修复从聊天界面改动记忆后,Memory 页面返回不刷新的问题
    • 修复占位目录影响记忆检索权重,以及超长网页内容直接塞进上下文的问题
    • 修复群聊、角色卡续聊、总结后自动续聊、实时渲染等多处稳定性问题

    优化

    • 对话历史和工具调用链路做了统一整理,复杂对话、工具调用和长上下文续聊场景更稳定
    • 本地模型的工具调用链路继续打磨,Llama 和 MNN 的行为更加一致
    • 消息现在可以记录更完整的模型、token 和耗时信息,单条消息详情更容易查看
    • 插件扩展能力继续完善,第三方功能更容易接入,也更容易拿到当前会话和调用来源信息
    • 深度搜索现在可以并发处理更多可执行任务,中间过程和总结输出也更清晰
    • 内置浏览器和网页工具继续打磨,复杂页面交互、局部快照、缩放和会话状态更稳定
    • 记忆系统加强了标签检索、重建进度展示和聊天界面同步
    • 包管理器扫描更快,能直接查看冲突来源,也支持删除外部冲突源文件
    • 大型文件修改结果和超长网页结果会更克制地返回给模型,减少无意义的上下文占用
    • 聊天渲染、浮动输入栏、历史过滤、批量删除聊天记录和整体浏览体验继续优化,整体更顺滑、可定制性更高
    • 原 all_about_myself 工具包正式更名为 operit_editor

    It adds a more complete in-app browser with browser-script support, liquid-glass theme effects and sidebar customization, FBX virtual avatars with an upgraded MMD preview experience, support for debugging and automatically writing Operit plugins with the configuration editor, workspace renaming, history jumping and bidirectional paging, plus more model, voice, image and workflow options.

    Added

    • Added FBX virtual avatar support and upgraded the MMD preview experience
    • Added a local HTTP chat entry point with health check, streaming callback support and optional floating-chat launch
    • Model configs can now define their own extra request headers
    • Character cards can now restrict available tools, and character cards/groups can keep their own custom emoji sets
    • Shared files and links can now be sent to the current chat, a recent chat or a new chat
    • Greatly expanded the in-app browser with tabs, history, bookmarks, permissions, multi-window support, minimization and viewport control
    • Added browser-script import, installation, storage, runtime and page-menu support for richer web automation
    • Added support for debugging and automatically writing Operit plugins with the configuration editor
    • Added page-by-page PDF vision parsing, app-usage-time reading and cold-start workflow triggers
    • Expanded automatic message context injection with current app info, current activity info, recent app usage and on-screen OCR text
    • Workspace root rule files can now be picked up automatically so project-specific rules can take effect more easily
    • Added workspace renaming, with linked chat titles updated together
    • Added fast chat-history scrolling, message jumping and bidirectional history paging
    • Added GIF support, HTML line breaks and more reliable formula rendering for richer message display
    • Added liquid-glass theme effects and more appearance options for bubbles, input bars and the sidebar
    • The sidebar now supports its own colors, highlight styles and animation options
    • Added more model and service choices including Novita, OpenRouter, MiniMax voice/image options and the

用户点评

登录用户的真实使用记录,可按版本和硬件筛选
共 0 条

    还没有人点评。第一个说出真实体验的人最有价值。

    README 原文

    未翻译

    🚀 Operit 2: Operit's Cross-Platform Successor

    This repository is Operit's Android edition. Operit 2 is a separate second-generation implementation centered on a shared Rust runtime, Flutter clients, and the operit2 CLI/TUI. It currently includes implementation or build paths for Android, iOS, Windows, macOS, Linux, and Web, with OpenHarmony support under active development. To follow the cross-platform version, visit Operit 2.

    Introduction

    Operit AI is an open-source AI Agent platform for Android that supports cloud and local models, connecting them to Android system capabilities, terminals, browsers, files, and project workspaces to perform real tasks such as information retrieval, file processing, code development, and device automation through tool calling, workflows, and extensions such as ToolPkg, MCP, and Skills.

    Highlights

    • A task-oriented agent workspace: Work with attachments, workspace context, tool execution progress, file changes, and multi-turn tasks
    • Your choice of models and services: Connect cloud models, custom compatible endpoints, and key pools, or use MNN models and GGUF models through llama.cpp
    • Deep Android and web integration: Use system capabilities after authorization, operate app interfaces, and read or interact with pages through the built-in browser
    • Mobile development and terminal environment: Manage projects, edit code, preview web apps, and use an Ubuntu 24.04 user space, SSH, and SFTP on your phone
    • Long-term memory and characters: Manage graph-based memory, chat history, character cards, and multi-character conversations, with an independent set of capabilities assigned to each character
    • Composable extensions and automation: Combine tools through the unified marketplace, workflows, and system integrations to create repeatable task flows

    Feature Showcase

    Main Features

    • Add images, audio, video, documents, and workspace files to conversation context
    • Use message branches, chat-history grouping and migration, automatic summaries, context limits, and parallel conversations
    • Connect OpenAI Chat/Responses, Anthropic, Gemini, and many compatible services, with additional providers available through ToolPkg
    • Manage multiple model configurations, parameters, key pools, and connection tests
    • Assign separate models to chat, memory, summarization, UI control, and other tasks
    • Run local inference with built-in MNN support or run GGUF models with llama.cpp
    • Connect to local model services such as Ollama and LM Studio
    • Use built-in tools for files, networking, search, media, system operations, software management, and development
    • Choose among Allow automatically, Ask every time, and Deny for each tool; the default setting asks for permission
    • Automate Android interfaces through Accessibility, ADB-level debugging access provided by Shizuku, or Root
    • Use PhoneAgent/AutoGLM with visual screen understanding to perform actions; capabilities such as virtual displays require the appropriate permissions and device support
    • Use the built-in browser with tabs, history, bookmarks, downloads, permissions, multiple windows, and user scripts
    • Let the browser agent inspect page structure, click elements, enter text, scroll, send key presses, and capture screenshots
    • Use OCR, image understanding, the camera, FFmpeg, web access, and file transfer tools
    • Start from Web, Android, Flutter, Node.js, TypeScript, Python, Java, Go, and other project templates
    • Use a file tree, code editor, syntax highlighting, live previews, change tracking, backups, and export tools
    • Bind a workspace to a chat so the AI can read project rules, reference files, and modify code
    • Access app-internal directories, SAF, SFTP, and SSH file systems as workspaces
    • Run an Ubuntu 24.04 ARM64 user space through PRoot by default, with chroot available on supported setups
    • Use multiple terminal sessions, Python, Node.js, vim, SSH, tmux, custom keys, and custom package sources
    • Work with development tools including Logcat, SQLite, Git, APKTool, and HTML packaging
    • Manage multiple memory spaces, imported and chunked documents, editable node relationships, and hybrid search
    • Extract information from conversations and attachments, then use temporal, semantic, and relational retrieval for long-term memory
    • Import, export, back up, and share character cards by QR code, including Tavern JSON/PNG formats
    • Bind separate models, memory spaces, tool packages, Skills, and MCP services to each character
    • Run multi-character group chats with mentions, separate histories, and collaboration between characters
    • Import, export, lock, branch, migrate in bulk, back up, and restore chat records
    • Search, install, and manage scripts, ToolPkg, Skills, and MCP services through one unified marketplace
    • Use the separate prompt and tag marketplace, along with project and Artifact management and publishing flows
    • Extend tools, interfaces, model providers, hooks, and runtime behavior through ToolPkg
    • Run local or remote MCP services, including launch methods such as uvx and npx
    • Build visual workflows from trigger, execution, condition, logic, and data extraction nodes
    • Trigger workflows manually, on schedules, through Tasker, intents, voice, or app startup
    • Inspect workflow logs and statistics, cancel runs, and manage workflows in batches
    • Use local Chinese and English speech recognition or cloud STT interfaces such as OpenAI and Deepgram
    • Use Android system TTS, local ONNX VITS, custom HTTP services, and multiple cloud TTS providers
    • Enable continuous voice conversations, background wake-up, custom wake templates, automatic read-aloud, and music queues
    • Open Operit from a floating window, chat bubble, home-screen widget, or the Android default assistant entry point
    • Use virtual avatars in DragonBones, WebP, MP4, MMD, glTF/GLB, and FBX formats
    • Customize themes, fonts, chat bubbles, backgrounds, toolbars, Markdown rendering, and layouts
    • Choose Chinese, English, Korean, Spanish, Malay, Indonesian, Brazilian Portuguese, or Romanian, or follow the system language
    • Optionally enable the LAN Web Chat and HTTP API, which are disabled by default; configure a bearer token and assess the risks of LAN exposure before enabling them

    Quick Start

    ItemDescription
    System RequirementsAndroid 8.0 (API 26) or newer; ARM64 (arm64-v8a) devices only
    Resource UsageMemory and storage usage depend on the terminal environment, installed tool packages, and local models; reserve space according to each model's documentation
    DownloadGet the latest APK from the Releases page
    User GuideVisit the Operit website for tutorials and examples

    Security notice: Only download installation packages from the official Releases page or the Operit website. Packages from unknown sources may be modified and may put your data or device at risk.

    Installation: Download the APK → Install and launch → Follow the setup flow to configure models and permissions → Start using Operit

    Data and network boundaries: Chats, characters, memories, and model configurations are stored locally by the app. Cloud-model requests are sent from your device to the provider endpoint you configure; Operit does not host chat inference. Marketplace, MCP, Skill, speech, and drawing features may connect to third-party services. Web Chat/HTTP API are disabled by default; review network exposure and bearer-token settings before enabling them, and use Android intent/broadcast integrations only with trusted apps.

    Project Evolution

    • April to August 2025 · From AI chat to tool execution: Added tool calling, MCP, voice, floating windows, web development, and character cards
    • September to December 2025 · Deeper device and development integration: Added the Ubuntu terminal, memory system, MNN, SSH workspaces, GUI automation, and the Skill ecosystem
    • 2026 to present · A mobile agent platform: Expanded workflows, GGUF local inference, the browser agent, character group chat, Web Chat, ToolPkg, and the unified marketplace
    • v1.12.0 · 2026-07-01: Introduced a unified marketplace and Artifact workflows, enhanced project workspaces and the ToolPkg runtime, expanded language, voice, and media support, and improved crash recovery
    • v1.11.0 · 2026-05-16: Added Web Chat, the Artifact marketplace, ToolPkg AI providers and hooks, and improvements to memory, context, and browser automation
    • v1.10.1 · 2026-04-17: Upgraded the built-in browser and web automation, added FBX avatars, liquid-glass themes, and a local HTTP chat entry point
    • v1.10.0 · 2026-03-18: Added character group chat, AI self-configuration, Ollama/NVIDIA/OpenAI Responses modes, and more development tools
    • v1.9.1 · 2026-02-20: Focused fixes for terminals, strict tool calling, remote MCP, memory, and workflows
    • v1.9.0 · 2026-02-17: Added mobile web automation, Windows terminal control, the SQLite viewer, and Android workspace templates
    • v1.8.1 · 2026-02-03: Added llama.cpp GGUF local inference and expanded interface, backup, Skill, and workspace capabilities
    • v1.8.0 · 2026-01-13: Introduced workflows, voice wake triggers, parallel conversations, and automatic backups
    • v1.7.1 · 2025-12-31: Added Root virtual displays, the Skill protocol and marketplace, region-based screen recognition, and chat locking
    • v1.7.0 · 2025-12-19: Introduced AutoGLM GUI automation and virtual displays
    • v1.6.4 · 2025-12-12: Integrated AutoGLM and improved automation status indicators, tool-package environment variables, and debug logs
    • v1.6.3 · 2025-12-07: Added native tool calling, multi-model configurations, SSH file systems, and workspace project templates
    • v1.6.2 · 2025-11-22: Expanded chat branching and migration, context binding, and academic search
    • v1.6.1 · 2025-11-16: Refactored UI rendering and added visual understanding, SSH terminals, automatic summaries, and Deep Search
    • v1.6.0 · 2025-10-21: Added MNN local models, intelligent memory, Tasker integration, and desktop pets
    • v1.5.2 · 2025-10-05: Improved MCP, workspace .gitignore support, camera tools, HTML rendering, and regex filters
    • v1.5.1 · 2025-09-25: Added the MCP marketplace, key pools, workspace rollback on chat resend, and manual memory refresh
    • v1.5.0 · 2025-09-20: Integrated the Ubuntu 24.04 terminal and Deep Search mode
    • v1.4.0 · 2025-09-01: Added parallel tool execution, the character card system, and PNG character card import
    • v1.3.0 · 2025-08-04: Added web development, the theme selector, and Anthropic Claude support
    • v1.2.x · 2025-07: Added voice conversations, the knowledge base, and DragonBones animation
    • v1.1.x · 2025-05 to 06: Added MCP, OCR, floating windows, and Gemini support
    • v1.0.0 · 2025-04-11: Released basic AI chat, tool calling, and Shizuku/Root integration

    See the Releases page for complete release notes.

    Open Source and Collaboration

    Contributions to Operit's scripts, extensions, documentation, and core features are welcome.

    Contributors

    Support Development

    If Operit is useful to you, you can voluntarily support ongoing development and basic project operations:

    • International support: Patreon
    • Mainland China: Afdian
    • Support is entirely voluntary and does not unlock features, quotas, updates, answers to questions, or other benefits
    • Choosing not to support does not affect normal use, updates, or access to the source code

    License

    The main code in this repository is licensed under GNU LGPL v3 (LGPL-3.0-only). Tools, examples, templates, and third-party dependencies may use other licenses; refer to LICENSE, the license files in their respective directories, and package metadata for their terms.

    Star History