The AI operating system where agents flow under pressure.
Browser-OS where memory, tools, recovery, and adaptation are wired in by default. Multi-agent orchestration as a substrate. Real APIs, no mocks. Agents build the apps live.
every line in that terminal hits real models. tokens, cost, ms — all measured live.
- · planner → executor → critic → memory
- · cockatiel retry + breaker fallback
- · Del Assistant · Gmail / Notion / GitHub / GDrive MCPs
- · voice autonomy via Whisper + 11labs
- · live STEER mid-stream
four power-ups, wired in.
Every DelOS run goes through all four layers, every time. Not an afterthought — the substrate.
Save State
MemoryHydraDB graph + vector recall across runs. Agents remember what worked, what didn't, what you told them last Tuesday.
Power-Ups
ToolsTyped tool registry with capability tags. MCP-shaped. Live marketplace with bundled + custom servers. Sibling fallback on failure.
1-Up
RecoveryCockatiel retry + circuit breakers. Critic agent triggers replans on drift. Model failover. Tool sibling fallback. Three layers of resilience.
Warp Zone
AdaptationGoal drift detection. User interrupts mid-stream rebuild the plan. Live STEER from any device. Context flood compressed at the boundary.
how it flows.
Every mission, every time. Deterministic loop, non-deterministic models.
Planner LLM reads goal + memory + decides fan-out. Outputs typed steps.
Executor picks a tool from merged registry (local + MCP) via cockatiel retry + breaker.
Critic scores drift after each step. Replans on >0.5. Mistral → Gemini failover.
HydraDB stores run summary by tenant. Local fallback keeps UI live during indexing.
one window, recursive.
A shared context window that spans message threads and the whole agent swarm. It remembers across chats, recalls on demand, and updates memory on every turn. Watch it live in the Context Inspector.
Every chat, run, and sub-agent is a node in one context tree. Children inherit the parent window; parents see what children learned. Depth is unbounded.
Ask once, recall everywhere. Search the whole swarm window across every thread and run you own — token-budgeted, ranked by relevance.
Each turn persists into the shared window automatically. Long-term HydraDB memory folds in on recall — no save button, no copy-paste.
Aged items fold into compact summaries so the window never overflows. The engine keeps the signal and drops the noise as context grows.
browser-OS where agents build the apps.
DelOS is a full window-managed desktop in a tab. Del Assistant (autonomous Gmail / Notion / GitHub / GDrive MCPs), Identity Core, Ingest Vault, VibeCode App Builder, DevFactory Cores — plus an App Builder that compiles your spoken prompt into a working mini-app live.
what you can actually do with it.
Six concrete patterns. Each one shippable today on free tiers. Each one tested live in the videos below.
Voice-only inbox triage
Say "summarize my last 10 emails and draft replies to the urgent ones." Reads Gmail, drafts in your tone, lands in Drafts ready to review.
Ship a SaaS prototype before lunch
"Build me a Stripe dashboard clone with KPI cards, charts, and a recent payments table." Real working app in 1.2s with full brand fidelity.
Ask Del Assistant to research anything
Open Del Assistant, ask "research the top 5 open-source AI agent frameworks". The assistant queries memory + web tools + drafts a markdown report you can paste into Notion in one click.
Survive provider outages mid-demo
Groq rate-limited? The orchestrator hops Mistral → Gemini → Bytez in under a second. The user never sees a 429. The chaos demo proves it on stage.
Cross-session recall via HydraDB
Yesterday you compared graph vs vector DBs. Today, ask Del Assistant. It remembers what you concluded, who you cited, and what you decided next.
Voice-drive the whole desktop
"Open terminal, calculate 17 times 19, then build me a habit tracker." Compound voice commands fan out into chained actions across the OS.
what you can build.
DelOS is the substrate. You decide the mission.
Research copilot
Multi-source web research with critic verification + memory across sessions.
Workflow automation
Chain agents, tools, MCPs into reusable missions. Voice-triggered.
Personal OS
Browser-based desktop where you build mini-apps on demand by talking to it.
Support agent
Del Assistant drafts Gmail replies, opens tickets in Notion, and references GitHub issues — all from one chat.
Data analyst
Plan → SQL → run → critique → re-plan if drift. Pipe to charts via spec apps.
Voice assistant
Whisper STT + ElevenLabs TTS + autonomy mode. Hands-free everything.
built on the right things.
No proprietary lock-in. Swap any provider via Settings.
questions.
Is this real or just a demo?+
Every endpoint hits real APIs: Groq Whisper for STT, ElevenLabs for premium TTS, Mistral / Gemini / Groq for LLM routing, HydraDB for memory, plus Gmail / Notion / GitHub / GDrive MCPs in Del Assistant. The token / cost counters in the Terminal aren't simulated.
Do I need API keys?+
Self-host with free-tier keys for Groq + Mistral + Gemini + HydraDB. ElevenLabs is optional — falls back to browser TTS. Total monthly cost at modest usage: $0.
What's MCP?+
Model Context Protocol — JSON-RPC tool servers. DelOS ships a bundled MCP server with 11 tools (crypto price, weather, dictionary, …) and accepts any HTTP MCP URL in Settings.
How does voice autonomy work?+
Hold the mic, say "build me a calculator," Whisper transcribes, a small LLM maps it to an intent action, the OS executes — opens App Builder, prefills the prompt, hits BUILD.
Can it run offline?+
The OS shell + games + window manager run offline. Agent calls need network. Install as PWA for offline shell. Tauri build available for native desktop.
How does cross-device sync work?+
Set the same Tenant ID in Settings → Device on every install. HydraDB partitions memory by tenant — instant sync across Mac / Win / Linux / iOS / Android / Chrome extension.
ship agents that
survive chaos.
Open DelOS. Speak to it. Watch four LLMs collaborate live. Build an app with your voice. Win.