2SinceJuly/SkyDot
AI EngineeringSelf-hosted multi-user agent workbench for creating named AI companions with shared group chats, scheduled tasks, skills, MCP, and memory.
Overview
SkyDot is an open-source, self-hosted multi-user agent workbench built on the Pi Agent Runtime. You create named, configurable agents (called Dots), converse with them individually, place several in the same group chat, and extend their long-running work through scheduled tasks, Skills, MCP, Plugins, and a memory system.
It is more than a chat UI. A workspace defines file and execution boundaries; the backend handles authentication, persistence, queues, scheduling, providers, channels, and authorization; and a Host or Docker Runner performs the actual agent execution. Tool approvals, device requests, and progress reports return to the same workbench.
The default interface centers on Dots, conversations, my work, and items requiring your attention, while workspace and runtime details live behind advanced entries. Web, PWA, and Electron share the same business frontend, and messaging channels such as Feishu and Telegram can be connected.
Quick Start
Prerequisites: Node.js 22.19+ (24 LTS recommended; the Pi Durable store needs Node's built-in SQLite) and npm. GNU Make and Docker are optional; a model provider is needed only for real AI conversations.
git clone https://github.com/2SinceJuly/SkyDot.git
cd SkyDot
npm install
npm --prefix web install
npm --prefix container/agent-runner install
npm run build:all
npm start
Open http://127.0.0.1:3000 and complete administrator initialization. Provider and channel setup can be deferred to settings; without a configured model, real AI conversations are unavailable.
For development, npm run dev:all starts the backend and Vite client together at http://127.0.0.1:5173. To browse the UI locally without login, use SKYDOT_DEMO_OPEN=true PROVIDER_SETUP_SKIPPED=true npm start — demo mode must not be exposed to the public internet or untrusted networks, as detailed in DEMO.md.
Container execution requires Docker and an agent runner image (default 2sincejuly/skydot-agent:latest); Host mode does not need Docker. Build the image from source with docker build --pull -t 2sincejuly/skydot-agent:latest container. The desktop app can be started with SKYDOT_RENDERER_URL=http://127.0.0.1:5173 npm run desktop:dev, and packaged with npm run desktop:package.
Constraints and Operations
The project is under active development; real models, Docker, external channels, and remote deployment each depend on their own credentials and runtime environments. Goal, Durable/Cloud Runtime, and device-node capabilities are gated behind switches and should be treated as experimental rather than default guarantees.
Several features are off by default: native multi-Dot group chat requires SKYDOT_BOT_GROUPS=true, bot execution collaboration additionally requires SKYDOT_BOT_HANDOFFS=true, and personal Goals require SKYDOT_PERSONAL_GOALS per the Goals and Tasks documentation. For third-party OpenAI Completions endpoints, set SKYDOT_MODEL_API=openai-completions in the provider's custom environment variables; the default protocol is anthropic-messages.
Runtime data defaults to the data/ directory (SQLite, workspaces, sessions, memory, and configuration), which is excluded from Git but contains sensitive material — protect backups the same way. Remote deployments should use HTTPS/WSS with proper reverse proxy, cookie, and access control; never expose the renderer, database, or Docker daemon to untrusted users. Permissions and security behavior are documented in the ACL matrix and security policy.
HotRepo's Take
SkyDot entered HotRepo's new-project table for TypeScript on 2026-11-16 with 48 stars gained over a window beginning 2026-11-15, ending at 183 stars and 16 forks. The hourly star samples show most of that growth concentrated on November 15, rising from roughly 135 to 182 within about eight hours; the earlier portion of the curve stayed near 120–135. With a single new-list record and no all-language daily ranking on file, the data supports a brief spike rather than a sustained trend. The technical offering — multi-user isolation, a clear backend/runner/workspace split, and a gated experimental feature set — is what readers can evaluate directly from the documentation.
License
MIT.
Written by AI from the project's README, updated . The project's own docs are the source of truth.
You might also like
- TencentCloud/OctopA self-hosted multi-user, multi-agent assistant platform for households and small teams that want one private instance.
- davidondrej/superbotSelf-hosted multi-agent workspace where named AI bots share one Linux VPS and run on coding-agent CLIs.
- yy838255939/duozhi-workbenchA local desktop workspace that puts several official AI websites on one screen and merges their replies into one group chat.