OpenWorkspace

A self-hosted AI workspace in a single Node process. Tasks, documents, calendar, scheduling, time tracking, a password vault, and a team of AI agents — behind one login, with no database to provision and no build step.

---

What is this?

OpenWorkspace is one Node process that hosts a suite of pluggable apps and AI agents for a small team — the kind of internal tooling companies usually stitch together from five SaaS subscriptions. It runs from a single npm start:

• No database — state is plain JSON / SQLite files next to the server.
• No build step — no webpack, no transpile, no framework. Open a file, edit, reload.
• Pluggable apps — every app is two files (<id>-api.js + apps/<id>.mount.js). Copy a template, you have a new app.
• A real agent layer — a chat assistant, a multi-agent framework you can scaffold new agents into, and an optional "AI OS" that gives the whole system presence, an event bus, and proactive insights.
• Built-in RAG — drop PDFs / Word / Excel into Documents and they're chunked, embedded, and searchable with hybrid keyword + vector retrieval (SQLite FTS5 + sqlite-vec). The agents answer grounded in your files — no external vector DB.

It scales from "a tasks app for my team" to "an AI-assisted operations workspace" without changing architecture.

Screenshots

Dashboard	Agents
Tasks	Scheduler

Quick start

git clone https://github.com/avaldesbosch/openworkspace.git
cd openworkspace
npm install
npm start            # http://localhost:8080  (set PORT to change)

Sign in with the seeded admin:

User	Password
admin	changeme

You'll set a new password on first login. That's it — Tasks, Documents, Calendar, Scheduler, Time Tracking and the rest work immediately. All state is written to gitignored files next to the server.

Apps

App	What it does
Tasks	Shared task tracker — assignees, priorities, due dates, projects, files, comments.
Documents	A full RAG engine, not just file storage. Upload PDFs / Word / Excel / images — text (and AI image descriptions) are extracted, chunked, and indexed for hybrid search: BM25 keyword (SQLite FTS5) + semantic vector search (sqlite-vec, 1024-dim embeddings) fused with reciprocal-rank. The agents answer questions grounded in your files. Per-user folder permissions throughout.
Calendar	Connect Microsoft 365 / Google calendars; free-busy lookups.
Scheduler	Cron-style recurring jobs that run prompts/agents inside the workspace.
Time Tracking	Log and report time; optional external time-system sync.
Passbolt	Credential vault (SSH keys, tokens, env bundles) the agents can use.
Arcade	A built-in game, because every workspace needs a break room.
Agents	Chat with the assistant roster (see below).
Command Center	Bug & idea tracker + triage board.
Admin	Users, app registry, AI model routing, agent scaffolding.

The agents

Agent	Role
Nexus	The always-on workspace assistant — reads across your apps and helps you act.
Forge	Ops / sysadmin agent — SSH, probes, log tails against a host allowlist.
Pulse	Scheduling agent.
Hunter	QA — drives the apps like a user and files bugs.
Builder	Writes docs + behavioral contracts for new agents.

To turn the AI features on, point a model at any OpenAI-compatible endpoint (OpenAI, Ollama, vLLM, LM Studio…). llm-config.json is auto-created from the example on first boot — edit it, or use Admin → AI Manager → Models to add one in the UI. Until then, the non-AI apps work fully and the AI surfaces show a clear "no model configured" state.

The agent framework

The agents above aren't hardcoded — they're built on a framework you can extend:

• Scaffold a new agent in ~a minute. Copy agent-templates/, give it a prompt and a tool list; it gets a full chat runtime, memory, and a registry entry. No restart.
• Tools come from "packs." Every capability an agent has (read a file, send a notification, run a query, SSH a host) lives in a composable pack under packs/. Grant packs per-agent.
• MCP built in. Point a pack at any Model Context Protocol server (stdio) and its tools surface automatically — bring any MCP tool into the workspace.
• Agents are cross-aware and delegate. Each agent's prompt lists the others; an orchestrator can hand a focused subtask to a specialist (ask_agent, one hop) and get a grounded answer back.
• Contracts enforced by a gate. Builder writes a behavioral contract for every agent/app, and lib/contract-coverage.js is a programmatic gate that rejects shallow contract sets — so coverage is measured, not vibes.
• Pick the model per agent. ai-router.js is one place to map "which model does this agent/role use right now?" — mix a cheap local model for background work with a frontier model for the assistant.

The AI OS (optional, opt-in)

lib/aios/ is an ambient intelligence layer that sits over everything — ~38 small modules, off by default, enrolled per app:

• An event bus every meaningful action publishes to (tool calls, state changes, cron ticks, user actions, agent observations).
• A condition registry + judge — rules describe what might be worth a user's attention; a judgment step decides what actually is, so it doesn't spam.
• Daily insight passes — once a day it reads the day's activity and surfaces what no rule would catch.
• Autonomy — it can propose actions and carry them out on your confirmation, not just observe.
• Per-user behavior model + weekly self-evolution — it learns each user's rhythm and proposes platform-level improvements to itself every week.

Hit /api/health to see the event bus live.

Voice (optional): browser mic → Whisper (STT) and text → Piper (TTS) via voice-api.js, so any agent can be talked to and talk back.

Architecture

server.js            boots apps, builds a prefix-sorted route table
  ├─ lib/router.js          the dispatcher
  ├─ apps/<id>.mount.js     route declarations (data)
  ├─ <id>-api.js            the app's HTTP handler + JSON state
  └─ public/<id>/           the app's frontend
packs/               composable tools the agents call (+ MCP servers via lib/mcp-bridge.js)
ai-router.js     which model each agent/role uses
lib/scaffolded-agent.js   runtime for scaffolded agents
lib/contract-coverage.js  the contract gate
doc-rag.js           hybrid BM25 + vector RAG (SQLite FTS5 + sqlite-vec)
lib/aios/            optional "AI OS": event bus, conditions, insights, autonomy
voice-api.js         optional STT (Whisper) + TTS (Piper)

Add an app in four steps — copy app-templates/, drop in two files, wire it into server.js, add a registry entry. The dispatcher sorts routes by prefix specificity, so order never matters.

Configuration

Env var	Default	Purpose
PORT	8080	HTTP port
COOKIE_NAME	workspace_session	session cookie name
PUSH_SUBJECT	mailto:admin@example.com	VAPID contact for web push

Optional integrations (Microsoft/Google OAuth, a Passbolt server, external time-tracking) are configured by an admin in-app and are off by default.

Data & secrets

Everything written at runtime — users.json, *-state.json, *.sqlite, data/ — is gitignored. The only secret file is llm-config.json (also gitignored; auto-created from llm-config.example.json on first boot). There are no credentials in this repository.

Contributing

Issues and PRs are welcome. The codebase is plain JavaScript with no build step, so you can clone, npm start, and start editing immediately.

License

MIT.