v1 release · the stack runs in production; UX and edges are still being refined

Multi-agent AI
that you run.

Pandora's Box runs multiple AI assistants simultaneously on one Mac or Linux mini-PC — one per company or context — with strict OS-level isolation, an independent oversight daemon, and a browser-first personal AI interface. Runs on your Claude Pro or Max subscription. Apache 2.0.

A Hope Inside Limited project · [ Pandora's Box ]
Open Source Multi-Agent Self-Hosted Apache 2.0
Capabilities

What the framework gives you

Security
Dependency scans · AI-independent oversight layers
Efficiency
Intent classifier on a local model · runs on Claude Pro / Max subscription
Memory
Vector recall · Obsidian vault · persistent across sessions
Self-Improvement
Shared improvement loop · skill governance · signed manifest
Scheduling
Cron and ad-hoc jobs · time-aware routing
Multi-Surface
Telegram · WhatsApp · browser · phone · watch · voice
Multi-Agent
Specialists in parallel · per-agent task isolation
Multi-Tenant
OS-level isolation · zero cross-tenant data bleed
Sub-Systems
Specialist subsystems · modular tools and skill library
Productivity
Mail · calendar · files · voice · marketing · web
Key features

What you actually get

From the AI Assistant you open in a browser tab to the Admin that runs every service underneath — every layer is here, designed to work together, owned by you.

AI Assistant with a full browser interface

The personal AI runs as a full Chrome app, not a chat box bolted onto a sidebar. Hold a voice-to-voice conversation, share your screen with it for live support, share your camera, let it see what you see, or hand it the wheel and let it drive a real browser session for tasks the chat can't do directly. Phone app and Pixel Watch surfaces for when you're away from the desk.

Chrome app Voice-to-voice Screen share Video share Remote browser control Phone app Pixel Watch

An Admin agent that runs the platform

A dedicated Admin oversees every service, runs scripted deploy sessions end-to-end, holds the systems dashboard, and operates a project + multi-project build system that can spawn its own sub-agents for parallel work. The platform tends itself, with you in the loop where it matters — and only where it matters.

Systems dashboard Scripted deploy sessions Project + sub-project system Parallel sub-agent spawning Service monitoring Browser terminal

Per-company agents, compartmentalised

Each company gets its own assistant with mail, calendar, files, voice, marketing and web surfaces — every one isolated at the operating-system level under its own user, with its own credentials and its own job queue. No cross-tenant data path. Telegram access is owner-allowlisted and role-gated per user before any message is even classified.

OS-level isolation Per-tenant credentials Six task agents per company Role-based access control Owner allowlist

Security that acts, not just watches

Oversight runs as its own process outside every agent and approves or blocks every queued action before it executes. A content classifier scores outbound text across six axes. A lockdown manager can quarantine an agent on demand — manually, after N strikes, or on a tamper trigger. A watchdog runs twice-daily integrity checks, a weekly dependency scan generates patch proposals for operator approval, and every important file backs up encrypted off-site each week.

Independent oversight Automatic lockdown Automatic quarantine Content classification Watchdog · twice-daily checks Weekly dependency scan Encrypted offsite backup

Memory built for continuity, with a self-improvement loop

Memory is structured into three working layers — a short-term rolling history of the current conversation, a long-term semantic vector recall over everything you've ever said to it, and a structured knowledge layer (knowledge vault, important standing facts, passive drops from other agents, a people-and-threads relationship graph). On top sits a six-worker self-improvement loop: reflection, calibration, prompt optimisation, dependency drift, a weekly digest to you, and nightly consolidation. Nothing self-applies without your approval.

Rolling history Semantic vector recall Knowledge vault Relationship graph Important standing facts Passive drops Six-worker self-improvement loop Weekly review

Specialised systems for real work

Beyond chat: an offline knowledge library carrying the full English Wikipedia and medical / survival / DIY / finance reference packs, answered locally with no internet required. A media production pipeline for PRS-free music, narration, image and video generation, and YouTube publishing. A trading-research agent for live market data and proposals. End-to-end app-publishing and book-publishing pipelines. Asynchronous deep research that returns a synthesised brief instead of a search results page. Sandboxed code execution. Camera and screen vision. Remote Windows diagnostics for B2B IT support. Intent classification on a local model so frontier reasoning is only spent where it earns its cost — with automatic fallback if the local model is unavailable.

Offline knowledge library Media production pipeline YouTube publishing Trading research Deep research Doc / image / music / video generation App + book publishing Sandboxed code execution Live-stream vision Windows remote diagnostics Local-model classification + fallback
Selected projects

Production-grade work on the platform

See all projects →

Architecture

Six layers, one direction of flow

Six layer architecture overview

Higher layers cannot be instructed by lower ones. Jobs flow down. Control flows up through the oversight daemon only. Read the architecture docs →