2026年10月30日October 30, 2026 · 产品Product

基于 DeepSeek Harness 的自托管团队工作台A self-hosted team workbench built on DeepSeek Harness

状态:建设中(公开开发)。main 分支 · package 0.1.0 · 更新至 2026年10月。P0–P3 已完成;P4 交付主体就绪;Docker 空卷一键部署已实机通过。Status: building in public. Branch main · package 0.1.0 · updated October 2026. P0–P3 done; P4 delivery largely ready; one-command Docker empty-volume deploy verified on a live machine.


基于 DeepSeek Harness 的自托管团队工作台。成员各自登录、各自独立工作区;模型的钥匙集中保管、按人计量。大家一起用 agent 干活,产出、账目、方法都留在自己的机器上。A self-hosted team workbench built around DeepSeek Harness. Each member signs in with their own account and works in an isolated workspace. The model API key stays on the server and usage is metered per person — so a team can share one agent setup without sharing keys, files, or bills.

一句话:引擎管干活,ANYWORK 管「谁能用、用在哪、花了多少」。引擎复用官方 DeepSeek Harness(MIT,不改源码);本仓库只做外壳——门户登录与反向代理、每人一个工作台实例、模型网关计量与预算。In one line: the engine does the work; ANYWORK decides who gets to use it, where, and at what cost. The engine is official DeepSeek Harness (MIT, unmodified). This repo is the shell — portal login and reverse proxy, one workbench instance per member, and a model gateway with metering and budgets.

ANYWORK workbench session
每位成员的工作台:浏览器即开即用,也可安装为独立应用窗口;界面中英双语。Each member's workbench — browser or desktop app window; bilingual 中文 / English UI.

门户登录 + 反向代理。统一闸门:未登录跳登录;登录后 HTTP 与 WebSocket 透传到该成员自己的工作台实例,跨成员访问拒绝。Portal login + reverse proxy. One gate: unauthenticated users hit login; after sign-in, HTTP and WebSocket traffic pass through to that member's own workbench instance — never someone else's.

每人一个独立工作台实例。独立 home、工作区与端口;systemd 自启守护;隔离复查脚本已绿。成员管理支持一键建号自动配齐钥匙、端口、插件与实例。One workbench instance per member. Separate home, workspace and port; systemd autostart; isolation checks green. Admins can add a member and have key, port, plugins and instance wired automatically.

模型网关:计量、预算、多通道。真 key 不下发;按人、按模型记账(峰谷价);月度预算 80%/100% 告警与超限拦截。DeepSeek、通义千问、Kimi、智谱 GLM、硅基流动、OpenRouter 等通道开箱可配,含单次输出上限防 400。Model gateway: metering, budgets, multi-channel. Real keys never leave the server; usage is billed per person and model (peak / off-peak). Monthly budgets warn at 80%/100% and block when exhausted. DeepSeek, Qwen, Kimi, GLM, SiliconFlow, OpenRouter and more — with per-channel max-output caps to avoid upstream 400s.

团队模块已就绪。任务板(提交 → 验收闭环,可关联会话审计);团队资料(公司盘文件 + 知识库笔记);通知与公告(12 种通道,含微信自助扫码、Telegram、WhatsApp 等);会话删除走管理员批准制。Team modules, ready. Task board with a submit → review/accept loop (session-linked for AI provenance); team library (shared drive + knowledge notes); notifications and announcements (12 channel kinds, including self-serve WeChat QR, Telegram, WhatsApp); session deletion with admin approval.

侧栏 WorkBuddy 式模块。助理(预设一览)、技能·连接器、自动化(定时任务卡片 + 快捷动作)、子代理(工作动态 / 定义管理)、Bots(常驻身份与群聊)、公告、团队资料、用量——点开即整页。WorkBuddy-style sidebar modules. Assistants (presets), Skills & Connectors, Automation (scheduled task cards + quick actions), Subagents (live activity / definitions), Bots (persistent personas and group chat), Announcements, Team library, Usage — each opens as a full page.

对话壳与近期打磨(2026-10)。聊天壳 UI 从工作台切出;插件库与预设模板重设;手机侧栏修复(点会话仍收栏、更多菜单不再误缩,#26 / #27)。Chat shell and recent cuts (Oct 2026). Chat-shell UI extracted from the workbench; plugins vault and presets-template redesign; mobile sidebar fixes (session tap still collapses; “more” no longer shrinks the rail — #26 / #27).

Usage panel — per-request cost tracking
用量、任务板与成员管理,都在同一个工作台里。Usage, task board and member management — all in one workbench.

路线与交付状态Roadmap and delivery status

P0–P3 已完成并在局域网实机验证(门户 / 网关 / 多实例 / systemd 自启)。P4 交付主体就绪:setup / backup / restore / logs 脚本、一页部署说明、端到端验收记录;Docker 一条命令空卷部署已实机全链路通过;第二台物理机从零重装待实测。P0–P3 are done and verified on the LAN (portal / gateway / multi-instance / systemd autostart). P4 delivery is largely ready: setup / backup / restore / logs scripts, a one-page deploy guide, and a green end-to-end acceptance walk; the one-command Docker empty-volume deploy completed a full from-scratch run; a second physical machine dry-run is still pending.

一些数字:每位成员 1 个独立实例;1 条 docker compose 命令部署;服务核心 0 个 npm 依赖(Node 24 直跑 TypeScript);界面 中 / 英 即时切换。A few numbers: 1 instance per member; 1 docker compose command to deploy; 0 npm dependencies in the core service (Node 24 runs TypeScript directly); 中 / English UI, switchable instantly.

快速开始:Quick start:

git clone https://github.com/bugbyc0922/ANYWORK.git && cd ANYWORK
cp .env.example .env # 填 ANYWORK_HOST 与 DEEPSEEK_API_KEY# set ANYWORK_HOST and DEEPSEEK_API_KEY
docker compose up -d --build

数据全在 Docker 卷里,升级/重建不丢。详见部署说明与使用手册;路线图见PLAN.md。All data lives in a Docker volume and survives upgrades. See the deployment guide and user guide; roadmap in PLAN.md.

已内置模型通道Built-in model providers
  • DeepSeek
  • 通义千问Qwen
  • Kimi
  • 智谱 GLMGLM
  • 硅基流动SiliconFlow
  • OpenRouter

1需要 Docker 与网络;首次构建约 10–25 分钟(含引擎编译),之后升级走缓存。详见部署说明。Requires Docker and network access. First build ~10–25 minutes (compiles the engine); later upgrades are cached. See the deployment guide.

关于 ToKenAIAbout ToKenAI

ToKenAI(PT TOKEN ARTIFICIAL INTELLIGENCE)是一家注册于印度尼西亚雅加达的人工智能公司,专注于 AI 系统的建设与运维——从模型接入、算力与成本治理,到让智能体真正在团队里干活的完整链路。我们相信,好的 AI 能力应当像水电一样:稳定、可控、留在最需要它的地方。ANYWORK,就是我们把这个想法产品化的第一步。ToKenAI (PT TOKEN ARTIFICIAL INTELLIGENCE) is an artificial intelligence company registered in Jakarta, Indonesia, focused on building and operating AI systems — from model access and compute-cost governance to the last mile, where agents get real work done inside teams. We believe good AI capability should be like a utility: dependable, controllable, and kept close to where it's needed. ANYWORK is our first step in productizing that belief.

注册地址Registered office

PT TOKEN ARTIFICIAL INTELLIGENCE

Gold Coast Office Tower Liberty Lantai 11 Unit H, Jalan Pantai Indah Kapuk Boulevard, Desa/Kelurahan Kamal Muara, Kec. Penjaringan, Kota Adm. Jakarta Utara, Provinsi DKI Jakarta, Kode Pos: 14470