Key Concepts

Before diving into specific features, it helps to understand the core concepts that power the platform. This page explains the mental model behind the product.

Accounts and Workspaces

Personal Account

When you sign up, you get a personal account — your private workspace where only you can access content. Use it for individual projects, experimentation, and personal AI workflows.

Team Accounts

Create or join team accounts to collaborate with others. Team accounts have:

• Members — People invited to the team
• Roles — Owner, Admin, and Member with different permissions
• Shared resources — AI agents, configurations, and usage metrics visible to the team
• Separate billing — Each team account has its own subscription

You can belong to multiple teams and switch between workspaces instantly.

AI Models and Providers

The platform connects to multiple AI providers, giving you access to different model families:

You choose which model to use for each conversation or task. Different models have different strengths, speeds, and costs.

Agents

An agent is an AI that can use tools to complete tasks — not just generate text. Agents can search the web, read files, execute code, create pull requests, and interact with external services.

Web Agents

Available on the web platform, these agents handle specific use cases:

• GenAI Studio — Multi-model chat interface
• Jira Triage — Ticket analysis and categorization
• Research Agent — Web search and analysis
• People Search — Contact research
• LLM Arena — Blind model comparison

Desktop Agent

The desktop app provides a general-purpose agent with deeper capabilities:

• Workspace access — Can read and write files in your designated directory
• Tool integration — Uses MCP servers for external services
• Memory — Remembers context across sessions
• Autonomous pipelines — Can run multi-step workflows unattended

Two-Phase Execution

The desktop agent uses a distinctive two-phase workflow:

Phase 1: Planning

When you give the agent a task, it first creates a plan:

• Analyzes your request
• Identifies what files, tools, and steps are needed
• Presents the plan for your review

Phase 2: Execution

After you review and approve the plan, the agent executes:

• Follows the planned steps
• Uses tools (file editing, code execution, API calls)
• Streams progress in real-time
• Produces results you can review

This approach gives you visibility and control — you always know what the agent will do before it does it.

Tools and MCP

Tools are capabilities that agents use to interact with the world. The platform uses the Model Context Protocol (MCP) — an open standard for connecting AI agents to tools.

Examples of tools:

• run_script — Execute code in a sandbox
• linear_create_issue — Create a Linear issue
• gmail_send — Send an email via Gmail
• memory_recall — Search long-term memory

Tools are provided by MCP servers — services that expose a set of related tools. You can install community MCP servers or create your own.

Sessions and Tasks

Sessions (Web)

A session is a conversation with an AI model or agent. Sessions preserve:

• Message history
• Context from previous messages
• Token usage and cost tracking

Tasks (Desktop)

A task is a unit of work for the desktop agent. Tasks have:

• A natural language prompt (your request)
• A plan (generated by the agent)
• Execution results (files created, commands run, etc.)
• Status tracking (planning, executing, completed, failed)

Tasks persist across app sessions — you can close and reopen the app without losing progress.

Memory

The desktop agent includes a long-term memory system that persists across sessions:

• Auto-recall — Before starting a task, the agent searches for relevant memories
• Auto-capture — After completing a task, important facts are saved automatically
• Manual storage — Tell the agent to "remember that..." for explicit storage
• Categories — Memories are categorized as preferences, facts, decisions, or entities

Memory uses a hybrid search combining vector similarity and full-text search for accurate recall.

Workspace

A workspace is the directory on your machine where the desktop agent operates. All file operations are confined to this directory for security.

Key workspace concepts:

• Isolation — The agent cannot access files outside the workspace
• Folder permissions — Each subdirectory requires explicit access consent
• Generated files — Files created by the agent are stored in the workspace
• Sandbox — Code execution runs within OS-level security boundaries

Skills

Skills are reusable instruction sets that specialize the desktop agent for particular tasks. They're defined as SKILL.md files with:

• A description of what the skill does
• Instructions for the agent to follow
• Tool requirements
• Example prompts

The Skills Marketplace offers 10,000+ community-created skills for common workflows.