openmeld

amplifthqMIT活跃维护

开源的人机协作系统,让团队成员和各类 AI 编程代理共享上下文、协调任务并并行工作,支持 Claude Code、Codex、Cursor 等。

The open-source collaboration system for people and AI agents. Share context, coordinate tasks, and work in parallel with your team, cloud agents, Codex, Claude Code, Grok, Grok Bot, OpenCode, Antigravity, and Cursor.

v0.3.263发布于 2026-10-16 10:27 UTC1,417star +4 / 7 天89fork同步于 2026-11-10 23:49 UTC

怎么上手

由 README 提炼,执行前请确认
npm install -g openmeld@latest
openmeld setup
硬件要求按 README 说明

README 未说明显存或 GPU 要求;本地代理需保持电脑在线连接,另提供免费云端计算机(不含免费模型用量与无限算力)。

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

开源人机协作系统,让团队与本地及云端 AI 代理共享上下文、并行推进任务

开源的人机协作系统,让团队成员和各类 AI 编程代理共享上下文、协调任务并并行工作,支持 Claude Code、Codex、Cursor 等。

核心能力

  • 在共享 Space 中接入 Codex、Claude Code、Cursor CLI 等本地代理
  • 由 Center Agent 托管 AI 队友协调任务并跨会话保留记忆
  • 任务可配置完成前审批,支持人与代理共同认领
  • 共享文件与文件夹并保留历史版本,支持定时运行任务
  • 通过公开链接分享 Space、会话、文件、Frame、页面或应用

容易踩的坑

npm 安装需 Node.js 22、24 或 25(推荐 24),23 及低于 22 会导致原生依赖安装失败,26 及以上不支持;Web 端连接本地代理会安装自带运行时的独立 CLI,无需升级 Node.js;本地代理运行期间电脑必须保持连接。

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

2 个版本v0.3.263
  1. v0.3.263最新正式版2026-10-16 · 2026-10-16 10:27 UTC

    v0.3.263

    Released October 16, 2026

    Improved

    Clearer setup guides and product answers

    The website now includes answers about working with existing agents and self-hosting, with links to the first-task guide and deployment limits. FAQ answers are available when the page first loads.

    The seven local-agent guides explain how to connect each tool and share task results. The Codex guide also links to access and tool-permission guidance.

    Find your Agents in the team directory

    Switch between My Agents and All Agents in Team's existing Agents page. Open an Agent's profile and settings, then return with your view, search, and filters preserved for the current workspace.

    Open Routines directly from the main navigation

    Routines has its own main navigation entry for personal and shared scheduled work. Team keeps People, Agents, Files, Skills, and Plugins together.

    Sidebars now match the page: People shows members, Agent details show the current Agent list, and Files shows file locations. Skills, Plugins, and Routines keep their own search and filters without an unrelated member list.

    Spaces and Team sidebars share compact headings and collapse controls. Choosing a file location returns to its root, and switching Agents preserves the return path to the Agent directory.

    Files keeps view, sort, filter, and search controls together on the left, with New on the right. File destinations appear directly below the sidebar heading.

    Spaces and People keep search in their sidebar options menu until needed, then focus the revealed field. People options group computer filters in a submenu and remove the five-minute activity filter. Sidebar icons and labels align with their shared headings. A small Close search icon clears and hides the field; Escape does the same and returns focus to the options button.

    Newly pinned conversations appear first in Pinned, keeping the existing order of other conversations. You can still drag pinned conversations to reorder them. Smaller pinned avatars and a two-to-five-column grid fit more conversations in the sidebar, including Spaces with multiple members.

    Search messages across Spaces with less waiting

    Message searches across many Spaces spend less time waiting for each conversation to respond. Web, Desktop, and Agent message reads use the same improved reader, with the existing access checks, result order, and complete pagination.

    Switch Spaces with less sidebar work

    Switching between Spaces keeps the conversation sidebar in place and avoids rebuilding unchanged rows. This reduces navigation delays in Web and Desktop, especially when your sidebar contains many conversations.

    Fixed

    Clearer Space addressing and task continuation

    The CLI catches repeated leading recipient names before sending. Continue or fork a task from its original request or an Agent reply with recorded task details. Reply threads remain a separate choice.

    When a selected task is unavailable, OpenMeld explains how to select another source or start a new task instead of reporting a model-provider outage.

    Keep delegated work connected to its Task

    Delegating an existing Space Task records its assignee and reviewer together with the work request. Agents receive the same Task through implementation, result submission, review, and requested changes. Task progress remains separate from execution status, and a missing reviewer is reported instead of silently dropping the requested review. Handoffs reach only their intended Agent, including in Spaces using listen-all, and unrelated Project or issue updates do not discard a pending work request.

    Keep full-width Space messages aligned

    Full-width Spaces keep message avatars and the composer aligned with the header, with consistent side spacing whether the history slider is visible or hidden. Touch layouts hide the history slider. Thread spacing is unchanged.

    See new Agents in the member picker immediately

    New Local and Remote Agents appear c

  2. v0.3.262OpenMeld 0.3.2622026-10-15 · 2026-10-15 04:43 UTC
    查看更新说明

    v0.3.262

    Released October 15, 2026

    Improved

    Faster release checks with visible stage timings

    Release checks avoid downloading unchanged deployment receipts and read independent workflow histories concurrently. Automatic continuation reports time spent checking status and running each component, so operators can locate delays without repeating a deployment. Fresh evidence and release checks remain required.

    CLI publication now requires the exact Windows helper used to prepare the release manifest. Automatic continuation reports a missing helper pin before deploying another selected component, instead of rebuilding different native bytes during publication.

    Fixed

    Keep renderer releases within their selected scope

    Website-only and Docs-only releases no longer require product database inspection credentials. Nested builds can reserve their actual memory budget without manually configuring an outer wrapper, provided current capacity and memory pressure allow it.

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

    Quick start

    1. Open OpenMeld and create a workspace or join your team.
    2. Open Center Agent, your hosted AI teammate.
    3. Give it a small task. No local installation is needed.

    Turn these notes into a three-item agenda: choose a launch date, test signup, and assign the announcement.

    Want to use your local agents too? On macOS, open Computers in OpenMeld Web and follow the connection steps on the computer that holds your work. The Web setup command installs a standalone CLI with its own runtime; no Node.js upgrade is needed.

    For Windows or a manual npm installation, use Node.js 22, 24, or 25. Node.js 24 is recommended.

    npm only warns about unsupported Node.js versions by default. On Node.js 23 or versions below 22, native dependency installation can fail. Select Node.js 24 before installing. Node.js 26 and later are also unsupported.

    npm install -g openmeld@latest
    openmeld setup
    

    Follow the setup prompts. Then create a Local Agent and choose its computer, agent tool, and project folder. Your computer must stay connected while it works. Keep your existing agent accounts and subscriptions.

    Full setup guide · CLI reference · Develop from source

    Features

    Bring your local agents into a shared Space. Ask Center Agent to coordinate the work, keep files and context with the conversation, and review results with your team.

    FeatureOpenMeldRaftMulticaPaperclipDustViktorOpenHands
    Run your local agentsYesYesYesYes??Yes
    Built-in team chatYesYesIssue threadsIssue threadsYesVia Slack / Teams?
    Tasks for people & agentsYesYesYesYesYesAgent tasksAgent tasks
    Require task approvalConfigurableNot enforcedNot enforcedConfigurableNo??
    Shared files & foldersYesAttachments / appIssue filesIssue filesYesChat filesWorkspace
    Public sharing linksSpacesArtifacts app??FramesPages & appsConversations
    Hosted cloud agentsYesOwn computeOwn computeCloud waitlistYesYesYes
    Free personal cloud computerYesNoNoNoStarter creditsStarter creditsFree cloud tier
    Team knowledge baseYesAgent filesSkills / projectsWiki pluginYesSkillsSkills
    Memory across chatsCenter AgentYesBy runtimeSkills / connectorsMemory toolYesSDK opt-in
    FeatureOpenMeldRaftMulticaPaperclipDustViktorOpenHands
    Work in your local foldersYesYesYesYesVia CLINoYes
    Saved instructions (Skills)YesYesYesYesYesYesYes
    Previous file versionsYesArtifacts appDeliverablesDocs / WikiFramesPages only?
    Run tasks on a scheduleYesManaged agentsYesYesYesYesYes
    Apps & pluginsPlugin catalogApp marketplaceChat / Git / MCPConnectors / pluginsConnectors / MCPNative / connectorsPlugins / MCP
    Core licenseMITFSL-1.1-ALv2Apache + termsMITMITProprietaryMIT

    Integration labels describe available connection methods, not comparable catalog sizes. Raft Artifacts requires an installed App. Multica's license adds conditions to Apache 2.0. Core licensing does not cover every third-party asset or hosted service; see License.

    Snapshot reviewed October 15, 2026, aligned with the website comparison. Yes = supported within the defined scope; No = not supported in that scope; ? = unverified. Text names the supported scope, setup, or availability limit. This compares documented capabilities, not measured speed or task success. See definitions and first-party sources.

    Public links can share a Space, conversation, file, Frame, page, or app; the table names which. Task approval means a required sign-off before completion, not a tool permission prompt. Memory and scheduling can depend on the selected agent.

    A free cloud computer does not include free model usage or unlimited compute. See the Cloud Computer guide.

    OpenMeld is in active development. Interfaces and commands can change. Self-hosting status describes the current deployment limits.

    Documentation

    Start hereExplore
    Get startedYour first workspace and useful reply
    Spaces and collaborationConversations, agents, projects, and files
    Team setupMembers, access, plugins, and shared knowledge
    HelpAccount, computer, and agent problems
    Developer guideLocal setup, repository map, and validation

    See the documentation index for all categories and architecture references.

    Contributing

    Bug reports, documentation fixes, and code contributions are welcome. Read CONTRIBUTING.md for your first change and pull request.

    Report a bug · Suggest a feature · Get help · Code of conduct · Report a security issue

    Credits

    Built with Fluid Functionalism and many other community projects. Thank you to their maintainers and contributors. See Credits and Third-party notices.

    License

    OpenMeld's original code uses the MIT License. Third-party code and assets retain their own terms. See Credits for attribution and Third-party notices for licenses and additional restrictions.