QUAN 2.0 documentation

Overview

QUAN 2.0 is an agentic AI operating system: a coordinated team of specialized AI agents with persistent memory, real tools and connectors, and approval-gated actions, working across web, a native Mac app, Telegram, and iMessage. It is an operator, not a chatbot. You give it a goal, and it reasons, plans, uses tools, verifies its own work, and returns a finished result, pausing for your approval on anything consequential.

QUAN 2.0 is proprietary and closed-source. Access is by request. It is built and operated by InterSpace Labs (legal entity: InterSpace Distribution), which runs its own company on it.

Key concepts

• Agent. A configured operator with its own system prompt, default model, and tool set. QUAN ships shared agents (General, Sales, CMS) and lets members create their own.
• Persistent memory. Typed, editable long-term memory of people, projects, and decisions, split into a shared house scope and per-member private scopes.
• Tools. Real actions an agent can take: email, calendar, web search, code execution, browser automation, image generation, and more. Every tool is scoped and, where consequential, approval-gated.
• Connectors (MCP). External systems plugged in over the Model Context Protocol (for example Notion or GitHub). Adding one is configuration, not code.
• Reasoning core. The plan, execute, and verify loop that runs on harder tasks, with a critic that checks the work before it is returned.
• Orchestration. Delegating a goal across multiple specialized agents in parallel and synthesizing their results, under a strict budget.
• Approval. A staged action that waits for a human tap before it executes. Anything touching money, mail, customer records, or the outside world is staged.

How it works

Give QUAN a goal and it runs the same loop every time, from understanding to delivery, pausing only where you want a say.

1. Understand. It interprets the request and recalls everything relevant from persistent memory.
2. Plan. On non-trivial work it drafts a plan, decomposing the goal into ordered steps. Plan mode lets you review and approve the plan first; say go to proceed.
3. Route. It selects a model for the task. Members route through Low, Medium, or Max and QUAN auto-selects the underlying model; admins route across every provider and can use several models at once.
4. Act. It calls real tools and connectors, runs code in an isolated sandbox, and drives a real browser when a system has no API.
5. Delegate. For bigger jobs it fans out to specialized agents in parallel and synthesizes their work, with a manager keeping the effort on budget.
6. Verify. An adversarial critic reviews the result, catches errors and gaps, and triggers a bounded correction. Background jobs loop until they converge.
7. Confirm. Consequential actions are staged for one tap. Nothing leaves the building without you (unless you turn on auto-accept for a conversation).
8. Remember. It writes the outcome back to memory you own and promotes what worked into reusable know-how.

Channels

The full agent stack runs server-side. Each surface is a way in, and your agents and memory follow you across all of them.

• Web. A fast, full-featured workspace in the browser.
• Mac app. A native, universal (Intel and Apple Silicon), signed and notarized desktop app that installs warning-free.
• Telegram. Message your agents and approve actions on the go.
• iMessage. Text QUAN like a colleague, right from Messages.

Models & routing

QUAN is model-agnostic. It routes across frontier APIs, your own Claude subscription, open and self-hosted models, and dedicated speech, vision, and image engines, reaching for the right model for each task.

Tiers (team members)

Invited members do not pick a raw model. They choose a tier, and QUAN auto-routes the actual model per task and, on Max, leans on multi-agent orchestration.

Providers (admins)

Administrators see the full catalog and can route across providers including Anthropic (Claude), OpenAI, Ollama, DeepSeek, Moonshot, Kimi, Qwen, Gemini, GLM, MiniMax, Mistral, and more, plus a Claude Pro or Max subscription and dedicated image models.

Bring your own, or use what is included

• Included. Every plan ships with an all-inclusive catalog of all major models, open-source and frontier alike, including Claude Opus, Sonnet, and Haiku, available under a single monthly subscription.
• Free with your own. Run your own self-hosted models, or plug in your own API keys, and QUAN is fully free to use, with no subscription.

Connection routes are configurable and updated on request.

Capabilities

• Persistent memory and learning across people, projects, and decisions.
• Multi-agent teams with manager, planner, and critic roles.
• Real tools and actions: email, calendar, invoicing, web search, code, browser, image generation.
• MCP connectors for Notion, GitHub, and any Model Context Protocol server.
• Autonomous workflows: one instruction triggers plan, act, verify, and self-correct.
• Sandboxed code execution in Python and JavaScript.
• Computer use: drives a real browser for apps with no API.
• Document intelligence over PDF, DOCX, XLSX, images, audio, and video.
• Knowledge and retrieval from a private, searchable knowledge base.
• Voice and transcription.
• Offensive and defensive security tooling, scope-gated and approval-bound.
• Runtime self-extension: writes and installs its own tools, skills, and connectors.
• Designs and builds UI with a living design system.

Multi-agent orchestration

For work that is too large for a single pass, QUAN coordinates a team of agents:

• Planner decomposes the goal into steps.
• Manager fans the steps out to specialized agents in parallel and synthesizes their results.
• Critic reviews the synthesized answer, looking for errors, unsupported claims, and missed requirements, then triggers a bounded correction.

Fan-out is capped by a budget (a maximum number of sub-agents and a concurrency limit) so parallelism never runs away. Delegation is one level deep and respects ownership: an agent can only delegate to agents the same member owns.

Memory

QUAN keeps a Claude-Code-style file memory: one typed fact per file, with an auto-generated index loaded each session.

• Scopes. A shared house memory available to every agent, plus a private memory per member and agent.
• Types. Facts are typed as user, feedback, project, or reference, so recall stays relevant.
• Yours to control. You can view, edit, and delete what QUAN remembers about your business, end to end.
• Learning loop. Outcomes feed a digest; repeated successes are promoted into reusable procedures, and contradictions are reconciled on a schedule.

Security & trust

QUAN is powerful by default and contained by design. The model is treated as untrusted; the guardrails are the product.

• Hard infrastructure boundary. The agent can never touch production databases or servers directly.
• Human-in-the-loop. Email, payments, and customer writes are staged for explicit approval.
• Sandboxed execution. Code and browser automation run isolated, behind an egress firewall, never near your data.
• Encrypted vault. Secrets are AES-256 encrypted; the model sees names, never values.
• Per-member scopes. Every teammate sees only their own agents and the tools an admin granted them.
• Full audit trail. Every action the agents take is recorded and reviewable.

Usage & limits

Team members have weekly usage limits, calculated the way Anthropic does it.

• Per-tier weekly caps. Low, Medium, and Max each have their own weekly allowance, with Max getting a smaller separate cap (like Opus).
• Block until reset. When a tier is spent, that tier is paused until the weekly reset. Lower tiers keep working in the meantime.
• Visible meter. The model picker shows how much of each tier you have used and when it resets.
• Bring your own to remove limits. Self-hosted models or your own API keys are not metered.

Admins are never limited, and they set each member's weekly caps in the Team panel.

Team & administration

Administrators manage the workspace from the Team panel:

• Invites. Add members by email; each gets their own login and private workspace.
• Tool grants. Each member has a tool pool (a ceiling). Their agents can only ever use tools from that pool, enforced server-side.
• Weekly limits. Set Low, Medium, and Max weekly caps per member.
• Roles. Promote a member to admin (full catalog, no limits) or demote back to member.
• Member-owned agents. Members create and edit only their own agents, and only with their granted tools.

Getting access

QUAN 2.0 is available by request. Tell us a little about you and we will reach out with access.

Request access →

Questions: eric@interspacemusic.com.

FAQ

Is QUAN 2.0 a chatbot?

No. A chatbot answers questions; QUAN carries the work. It remembers context, plans, uses real tools, verifies its output, and returns finished results.

Is QUAN 2.0 open source?

No. QUAN 2.0 is proprietary and closed-source. Access is by request.

Which models can it use?

Members pick a Low, Medium, or Max tier and QUAN auto-routes. Admins can use the full catalog across many providers, a Claude subscription, local and self-hosted models, and dedicated speech, vision, and image models.

Can it touch my database or servers?

No. There is a hard boundary: the model has no direct access to production databases or servers, and code runs sandboxed behind an egress firewall.

What happens when I hit my weekly limit?

That tier pauses until the weekly reset, while lower tiers keep working. Using your own models or API keys removes the limit.

Glossary

• Operator. An agent that does the work, as opposed to a chatbot that only talks about it.
• Tier. A member-facing routing level (Low, Medium, Max) that maps to one or more underlying models.
• MCP. Model Context Protocol, the open standard QUAN uses to connect external tools and data.
• Approval. A staged consequential action awaiting a human tap.
• Plan mode. A mode where QUAN presents a plan and waits for your go-ahead before acting.
• Auto-accept. A per-conversation mode where staged actions run automatically.
• House memory. Shared memory available to every agent; private memory is scoped per member and agent.