teloa/teloa repository overview

⁠Teloa

AI-Native Team Studio. Build and run your own AI team.

Self-hosted Community edition · 0.2.8 · Linux ARM64

⁠Docker 安装 / Quick start

mkdir teloa-community
cd teloa-community
curl --fail --location https://docs.teloa.ai/downloads/0.2.8/compose.yaml --output compose.yaml
docker compose pull
docker compose up -d --no-build
docker compose ps
docker compose logs --tail=50 app

等待 db、app 健康后，从本人终端的 app 日志打开完整认证链接；一次性 init 服务成功退出是正常状态。3100 已占用时，启动前在此目录的 .env 设置 TELOA_PORT=3101。模型服务与密钥需自备。已有部署升级前先备份，保留原 Compose 项目与数据卷。容器中的“本地”是容器或部署服务器，不会自动取得浏览器所在电脑的文件和程序。

Wait for db and app to become healthy, then open the complete authentication link from your own app log. The one-time init service exits successfully. If port 3100 is occupied, set TELOA_PORT=3101 in this directory's .env before starting. Bring your own model service and credentials. Back up existing deployments and retain their Compose project and volumes when upgrading. “Local” inside the container means the deployment host, not the computer running your browser.

teloa/teloa:0.2.8 与 latest 提供同一镜像。Linux amd64 尚未验证。 / 0.2.8 and latest provide the same image. Linux amd64 has not been verified.

安装与升级⁠ · 完整更新日志⁠ · Install and upgrade⁠ · Full changelog⁠ · GitHub⁠

⁠简体中文

⁠AI 员工：持续负责一类工作

• 用一段 Markdown 描述岗位与工作要求，自动带入当前业务并保留草稿；核对生成的职责与边界后，再明确确认创建。
• 为员工设置岗位职责、工作方式、业务范围、资料、技能和可用工具；这些配置随员工保留，适合持续处理同类工作。
• 从员工目录进入私聊、查看任务、成果和经验。会话、员工和群分开管理，员工详情按紧凑总览组织，具体内容按需展开。
• 员工只在授予的业务与工具范围内工作；添加员工、配置能力和真正开始执行是分开的动作。

⁠AI 团队：把多个岗位组合成业务分工

• 市场提供独立的 AI 团队入口。团队模板可以同时包含多个员工、技能、参考资料、可选连接和工作模板，减少逐个配置的重复操作。
• 添加后核对员工与目标业务，再按实际需要启用连接和工作安排；模板不会自动接通账号、扩大权限或启动后台工作。
• 在群中与多位员工协作，使用明确的 @ 指定参与者，保留交办来源和结果回帖。跨业务团队沿用各员工的实际业务范围，并不因为同群就获得全部资料。
• 长期团队与临时执行小组分开：AI 员工是长期岗位；执行小组是当前任务内的临时分工，可以并行处理独立部分，再由负责者汇总，不会自动新增员工或加入业务群。

⁠多 Harness 执行：DSH、Codex 与 Claude Code

• 用自然语言指定本次任务，例如“这次用 Codex 检查代码”“让 Claude Agent 分析这些文件”，或要求自动选择可用引擎。本次选择不改永久偏好，也不自动更换聊天模型。
• 引擎不等于模型。 接入 Codex 或 Claude Agent，不代表自动获得某家模型服务、订阅或额度；按所选提供方准备对应认证与模型连接。普通会话、群与员工仍遵循原有工作流程。
• 外部引擎使用其原生权限，启动前按次取得本人同意；Teloa 的文件范围和逐项工具审批不会自动传入外部引擎，确认前需核对其实际工作目录与权限。
• 查看真实运行状态、处理所需授权，并把交付结果接回原任务。指定引擎不可用时保留原任务并说明原因，不悄悄改由另一个引擎执行。

⁠项目与项目记忆：让后续工作带着上下文

• 项目归集目标、简报、任务、员工、群、资料、自动化和成果，并支持专属项目会话。
• 项目目标、有效规则、相关资料和记忆可以进入后续模型请求；可查看「本次使用」及来源，而不只是保存一张关联清单。
• 审阅记忆候选、查看来源、撤回不适用的内容；新会话可选择不读写长期记忆，个人偏好是否参与项目由本人决定。
• 项目记忆分享固定在明确修订上，可撤回后续使用；撤回不能收回已经发送到外部模型的数据。

⁠资料索引与本地模型：检索和生成分别配置

• 明确选择本机 Markdown 文件，跟踪版本、加入索引，并查看检索引用及原文摘录位置；不自动扫描整台电脑。
• 在资料页的检索模型面板准备内置的 Qwen3 Embedding 0.6B 或 EmbeddingGemma 2，下载前确认，准备进度可见；检索模型列表与市场模型目录分开。
• 通过 Ollama 检查本地聊天环境、接入已有模型、确认下载或取消准备。聊天模型负责生成，嵌入模型负责检索，二者分别管理。
• 索引运行在部署 Teloa 的设备或容器上；若使用远程聊天模型，检索摘录仍可能发送给该服务。文件预览支持的格式多于当前本地索引支持的格式。

⁠会话与群聊：主消息流、话题和交办各归其位

• 普通会话、员工私聊和群分别列出；保留各入口的草稿和当前工作上下文。
• 群主消息的回复留在主消息流，并可回到来源消息；话题中的回复继续留在话题，避免所有回答堆到一个话题中。
• 明确的 @ 会随交办传递，执行时继续检查成员、业务、资料和工具授权；群内发言不代替本人审批。

⁠自然语言任务与版本化成果

• 主要任务入口使用一段 Markdown 说明工作，自动带入来源会话、项目或业务；保存草稿与开始执行分开。
• 在同一工作路径跟进执行、处理审批、停止工作、查看文件和取得成果。小松鼠及通用 Agent 在没有已创建员工时也可以独立工作。
• 预览和下载成果、按选段或整体反馈续改，保留 v1、v2 等历史版本及原归属；业务验收继续指向当时采用的明确版本。
• 已确认的方法可复用为技能或计划草稿；新计划在明确启用前保持暂停。员工主创建使用 Markdown 草稿；部分群配置仍使用结构化控件。

⁠业务记录、看板与应用市场

• 用自然语言创建或调整业务配置，审阅草稿后采用；导入记录、连接已授权来源，把任务推进到可验收成果。
• 「业务内容」与「统计看板」各显示自己的页面并记住选择；减少重复标题和说明，短卡随内容收紧，图表显示类别与数值，缺失数据不混作零。
• 市场以 AI 员工、AI 团队等用户可理解的入口组织发现与添加；应用页管理已安装能力与连接。模型、应用和业务授权各有对应配置，不因添加模板自动联动。

⁠语音输入、TTS 与通话

• 与员工或小松鼠通话，使用静音、打断回答和挂断；导航时保留可达的通话控制。关闭详情不会挂断，挂断也不停止原会话中已经开始的工作。
• 小松鼠支持语音留言，转写后进入原会话草稿；隐藏或收起助手不结束会话。
• 语音凭据与聊天模型凭据分别管理。

⁠Auto Dream、员工经验与个人分身

• Auto Dream 从工作证据整理私有每日小结，保存带来源和范围的员工经验，并提供查看与撤回。
• 员工经验、项目记忆与个人分身资料分别管理。分身的判断样本和习惯观察属于私有材料，不会自动注入每次普通对话，也不是模型权重训练。

⁠文件、工具与升级连续性

• 工作台预览 Markdown、代码、图片、PDF、HTML、Office 文件与表格，查看本轮文件及可用的文本改动，再按需保存明确版本的成果。
• 在授权范围使用 Workflow 分步编排、PTC 工具组合，以及本人明确要求的 Ralph 迭代；临时分工仍受本次工作范围和额度约束。
• Browser Use、Computer Use、IM 通道等可选能力分别配置、启用和授权，不因安装包中存在就获得设备控制权。
• 升级沿用原安装身份、数据和工作目录，切换前备份；减少重复文案，统一详情抽屉、导航、等待与失败状态，保留历史及未提交草稿。

⁠0.2.8 更新

• 员工 API 执行控制：按请求指定 danger-full-access / workspace-write / read-only 权限模式与自定义模型 provider，续办可显式继承或覆盖，并持久化到工作与轮次。
• API 结果自动交付：固定成功结果在同一事务生成正式成果、完成关联任务并更新工作进展；失败回滚，旧的成功未交付工作可直接补齐，不重新调用模型。
• 员工包资料闭环：资料正文可预览、员工单聊按岗位范围检索，资料库可按员工一键加入本地检索；更新与卸载继续保留身份、历史、共享资源和事务切换保护。
• 市场与预设：创建菜单补齐技能/应用；Teloa 标准预设归入内置，PTC 代码模式名称与说明修正；设置分类层级与非中文侧栏布局修正。
• 安全基线：容器运行时 Node.js 24.22.1。SDK 继续兼容 @teloa/[email protected]。

⁠English

⁠AI employees: lasting responsibility for recurring work

• Describe the role and requirements in Markdown, carry in the current business and preserve the draft. Review the proposed responsibilities and boundaries before explicitly confirming creation.
• Keep an employee's responsibilities, working methods, business scope, knowledge, skills and permitted tools together across tasks.
• Open employee conversations, tasks, deliverables and experience from the employee directory. Chats, employees and groups have separate lists; compact employee overviews reveal details on demand.
• Adding an employee, configuring capabilities and starting execution remain separate actions. Employees work within their actual business and tool permissions.

⁠AI teams: combine roles into a business workflow

• Discover AI teams separately in the marketplace. Team templates can include multiple employees, skills, reference materials, optional connections and work templates, reducing repetitive setup.
• Review the employees and target business, then enable the connections and work arrangements you need. A template does not connect accounts, expand permissions or start background work automatically.
• Collaborate with several employees in a group, use explicit @ mentions, and retain the assignment source and result replies. Sharing a group does not grant access to every member's business materials.
• Lasting teams and temporary execution teams are different: employees are persistent roles; execution assistants split up the current task, can work on independent parts in parallel, and return results for consolidation. They do not automatically become employees or group members.

⁠Multiple execution harnesses: DSH, Codex and Claude Code

• Specify the engine in natural language: “Use Codex to review this code,” “Have Claude Agent analyze these files,” or ask for an available engine to be selected automatically. A task-specific choice does not change permanent preferences or silently replace the chat model.
• An engine is not a model service. Integration does not include a provider account, subscription or inference allowance. Configure the authentication and model connection required by the chosen provider. Ordinary chats, groups and employees retain their working rules.
• External engines use their native permissions and require the owner's consent for each launch. Teloa's file scope and per-tool approvals are not automatically passed into the external engine; review its actual working directory and permissions before consenting.
• Follow actual execution and required authorization, then receive the result in the original task. If a requested engine is unavailable, Teloa preserves the task and reports the reason instead of silently substituting another engine.

⁠Projects and project memory: context for subsequent work

• Organize project goals, briefs, tasks, employees, groups, materials, automations and deliverables, with dedicated project conversations.
• Include the project goal, active rules, relevant materials and memories in subsequent model requests. Inspect what was used and its sources, beyond a list of linked objects.
• Review memory candidates and sources, withdraw unsuitable content, start a conversation without long-term memory, and explicitly choose whether personal preferences apply.
• Share a fixed memory revision and withdraw subsequent use. Withdrawal cannot recall information already sent to an external model.

⁠Document indexes and local models: retrieval and generation configured separately

• Select local Markdown files explicitly, track versions, add them to an index and inspect citations and excerpt locations. Teloa does not scan the entire computer automatically.
• Prepare the built-in Qwen3 Embedding 0.6B or EmbeddingGemma 2 in the Knowledge retrieval-model panel, with download confirmation and visible progress. This list is separate from the marketplace model catalog.
• Use Ollama to inspect the local chat runtime, connect an existing model, confirm downloads or cancel preparation. Chat models generate answers; embedding models support retrieval. They are managed separately.
• Indexing runs on the deployment device or container. Retrieved excerpts may still reach a remote chat provider. File preview supports more formats than the current local index.

⁠Chats and groups: distinguish stream replies, topics and assignments

• Keep ordinary chats, employee conversations and groups in separate lists, preserving drafts and working context.
• Replies to a group-stream message stay in the stream with their source; topic replies stay inside the topic rather than turning every response into a topic.
• Explicit @ mentions carry through to assigned work. Execution still checks membership, business scope, materials and tool permissions; a group message does not replace the owner's approval.

⁠Natural-language tasks and versioned deliverables

• Describe work in Markdown in the main task entry points, carrying the source conversation, project or business into the draft. Saving is separate from execution.
• Follow execution, handle approvals, stop work, inspect files and obtain deliverables in one working path. The companion and general Agent can work without any created employees.
• Preview and download results, revise a selection or the whole deliverable from feedback, and retain v1, v2 and later versions with their original attribution. Business acceptance continues to reference its selected version.
• Reuse an accepted method as a skill or plan draft. A new plan stays paused until explicitly enabled. Employee creation uses a Markdown draft; some group configuration remains structured.

⁠Business records, dashboards and apps

• Describe a business configuration or change in natural language, review and adopt its draft, import records, connect approved sources and progress to an accepted result.
• Records and Dashboards show their own pages and remember their selections. Cards remove repeated headings and explanations, fit their content, and show chart labels and values; missing data remains distinct from zero.
• Discover employees and teams in the marketplace; manage installed capabilities and connections in Apps. Model setup, application setup and business permissions remain explicit rather than changing automatically when a template is added.

⁠Voice input, TTS and calls

• Call an employee or the companion, mute, interrupt a response or hang up, and keep call controls accessible during navigation. Closing details does not hang up; hanging up does not stop work already running in the conversation.
• Record a companion voice message and review its transcription in the original conversation draft. Hiding the companion does not end the conversation.
• Voice credentials are separate from chat-model credentials.

⁠Auto Dream, employee experience and a personal twin

• Auto Dream organizes private daily summaries from work evidence and saves employee experience with sources and scope, with inspection and withdrawal controls.
• Employee experience, project memory and personal twin materials remain distinct. Judgment samples and habit observations are private materials, not automatic context for every ordinary conversation or model-weight training.

⁠Files, tools and upgrade continuity

• Preview Markdown, code, images, PDF, HTML, Office files and tables, inspect current-turn files and available text changes, and save a specific deliverable version when needed.
• Use Workflow steps, PTC tool composition and explicitly requested Ralph iteration within their approval boundaries. Temporary delegation retains the current work scope and allowance.
• Configure and authorize optional Browser Use, Computer Use and IM channels separately. Their presence in the package does not grant device access.
• Preserve the original installation identity, data and workspace during upgrades, with a backup before switching. More consistent drawers, navigation, pending/failure states and reduced repeated copy retain history and unsubmitted drafts.

⁠Updates in 0.2.8

• Employee API execution controls: per-request danger-full-access / workspace-write / read-only permission mode and custom model provider; continuations inherit or override explicitly, persisted on the work and round.
• Automatic API result delivery: a fixed successful result creates a formal artifact, completes the linked task and updates work progress in one transaction; failed delivery rolls back, and legacy succeeded-but-undelivered work can be repaired without recalling the model.
• Employee package materials: material bodies can be previewed, employee chats retrieve within the role scope, and the library can add all of an employee's materials to local retrieval in one action; update/unload keeps identity, history, shared-resource protection and transactional switching.
• Marketplace and presets: Create menu adds skill/app entries; the Teloa standard preset is grouped under built-ins, PTC code modes have clear names and accurate descriptions; settings hierarchy and non-Chinese sidebar layout are fixed.
• Security baseline: container runtime is Node.js 24.22.1. SDK remains compatible with @teloa/[email protected].