Core Concepts

Before diving deeper, here are the key concepts you'll encounter throughout Cosmo. Understanding these will make everything else click.

Data Hierarchy

Cosmo's data model follows a strict hierarchy. Every piece of data belongs to exactly one level:

Organization (tenant boundary)
  └── Team (access boundary)
        └── Workspace (primary work container)
              ├── Artifact (document / report / output)
              └── Memory (AI context files)

Organizations

An organization is the top-level tenant boundary. It owns:

• Billing and subscription
• All members and their roles
• All MCP integration configurations
• All teams, workspaces, and their data

Everything in Cosmo is scoped to an organization. When you sign up, you create one. You can belong to multiple organizations (e.g., your company's org and a personal org).

Examples: "Acme Corp", "Parag's Workspace", "Product Team"

Teams

A team is the access boundary within an organization. Teams control who can see and contribute to workspaces.

Every user automatically gets a personal team — a private space for drafts, notes, and work-in-progress that only they can see. Shared teams are created for collaborative work.

Teams are not organizational hierarchy (there are no "sub-teams") — they're a flat access control layer. A workspace belongs to one team. Members of that team can access the workspace.

Examples: "Product", "Engineering", "Sales", "Marketing", or a personal team like "Parag"

Workspaces

A workspace is the primary user-created container for work. It scopes:

• Artifacts — documents, reports, and other AI-generated outputs
• Memory — AI context that persists across conversations
• Integrations — which connected tools are active in this context
• Skills — which AI capabilities are available

Think of a workspace as a dedicated space for a specific initiative, project, or workstream. It gives AI enough context to be genuinely useful without mixing up context between unrelated work.

Examples: "Q3 Product Planning", "Customer Onboarding Redesign", "Sales Collateral", "Engineering Runbook"

Artifacts

Artifacts are the structured outputs AI creates for you — documents, proposals, reports, analyses, visualizations, data tables, and more. They're different from chat responses because they:

• Persist — they live in your workspace, not in a chat transcript that scrolls away
• Are versioned — every edit is tracked; you can review history
• Are collaborative — your team can see and build on them in real time
• Live in cloud storage — not in the database; they're actual files (HTML, Markdown, JSON)
• Are displayed in the Hub — the right panel renders them rich, not as raw text

You create artifacts through conversation. AI decides when to produce an artifact vs. a plain text response. You can ask explicitly:

"Create an artifact for this" or "Write this up as a document"

Artifacts can contain formatted text, charts (via Chart.js), tables, embedded data, and more. They can be moved between workspaces, exported, or shared with teammates.

The Hub

The Hub is the right panel in Cosmo's three-panel layout:

┌───────────┬────────────────────┬────────────────┐
│  Sidebar  │    Chat Panel      │   Hub Panel    │
│           │                    │                │
│ Sessions  │  ← conversation →  │  Artifacts     │
│ Dashboard │                    │  Tasks         │
│ Profile   │                    │  Workspaces    │
│           │                    │  Connectors    │
└───────────┴────────────────────┴────────────────┘

The Hub is your workspace canvas — it shows:

• Artifacts being generated in real time, fully formatted
• Dashboard with an overview of your org's activity
• Tasks across workspaces
• Workspaces browser
• Connectors for managing integrations

When AI generates an artifact, it appears in the Hub immediately as it streams, rendered with full formatting. The Hub is where Cosmo's outputs live — it's not just a viewer, it's the persistent record of what's been created.

Memory

Memory is how Cosmo gets smarter over time. Rather than starting every conversation from scratch, AI maintains context across sessions.

Memory is stored as plain files in cloud storage — not in a hidden vector database. You can read, edit, and correct memory files directly. The AI also reads and writes them during conversations.

There are three levels of memory:

Workspace Memory

Scoped to a single workspace. Contains:

• Key decisions made in this project
• Project-specific terminology and conventions
• Context about stakeholders, goals, and constraints
• Ongoing items and next steps

Location: /{orgId}/workspaces/{workspaceId}/memory/

Team Memory

Scoped to a team. Shared across all workspaces in that team. Contains:

• Team-wide conventions and working agreements
• Shared knowledge that applies to multiple projects
• Team preferences for how AI should behave

Location: /{orgId}/teams/{teamId}/memory/

Hub Memory (Org Memory)

Scoped to the entire organization. Cross-workspace patterns. Contains:

• Org-wide brand and tone guidelines
• Company conventions that apply everywhere
• Cross-workspace patterns AI has detected

Location: /{orgId}/shared/hub-memory/

How Memory Works in Practice

When you start a conversation:

1. AI loads relevant memory files for your active workspace, team, and org
2. That context shapes every response — AI already "knows" your conventions
3. When you share new context, AI may update memory files accordingly
4. You can ask AI to remember something explicitly: "Remember that we call our users 'members' not 'customers'"

MCP Integrations

Cosmo connects to external tools through MCP (Model Context Protocol) — an open standard developed by Anthropic for giving AI structured access to external services.

Each integration provides a set of tools — named capabilities the AI can invoke. For example, connecting Google Drive provides tools like search_drive, read_file, and create_document. When you ask AI to find something in Drive, it calls these tools directly.

This is fundamentally different from copy-pasting. The AI can:

• Read data from your tools (search, retrieve, list)
• Write to your tools (create tickets, send messages, update records)
• Take actions (run workflows, trigger automations)

Integrations are configured at the organization level (who can use them and their credentials) and enabled at the workspace level (which tools are active in a given context).

See Available Integrations for the full catalog.

AI Providers and Models

Cosmo supports multiple AI providers, which you can switch between at any time:

Different models have different strengths, speeds, and costs. A typical workflow:

• Use a fast, balanced model (Claude Sonnet or GPT-4o) for everyday work
• Switch to a more capable model (Claude Opus or o1) for deep analysis
• Use a faster/cheaper model (Claude Haiku) for bulk tasks

Enterprise plans can bring their own API keys (BYO keys), so AI usage is billed directly to their accounts.

Skills

Skills are pre-built AI capabilities for structured workflows. They're prompt packages — structured instructions that tell AI how to tackle specific tasks.

Built-in skills include things like:

• Executive report generation
• Project status documents
• Client proposals
• One-pager briefs

Skills appear as options when you start a conversation or can be triggered by name. They encode best practices for common outputs, so you get consistent, high-quality results without prompting from scratch each time.

Agents

Agents are autonomous AI workflows that execute multi-step tasks without step-by-step supervision. An agent can:

• Use multiple tools in sequence
• Produce multiple artifacts
• Access workspace memory and integrations
• Make decisions at each step based on prior results

Agents are different from a regular chat: in chat, you guide each step. With agents, you describe a goal and the agent executes a plan.

Example: "Research our top 5 competitors, find their pricing pages, and produce a comparison matrix artifact." — an agent would search the web, retrieve competitor data, synthesize it, and write the artifact, all autonomously.

Sessions

A session is a single conversation thread. Sessions are:

• Saved automatically with AI-generated titles
• Searchable from the sidebar
• Scoped to an active workspace (if set)
• Persistent — you can return to any past session

Sessions are not workspaces. A workspace is a persistent container for artifacts and memory. A session is a conversation — it lives within a workspace context, but many sessions can exist for the same workspace.

What's Next