Docs

Core Concepts

Understanding a handful of concepts will help you get the most out of Agentastic. Here's the mental model.

The launcher and the agent

The launcher is the window you summon with a hotkey. It's how you talk to the agent, the AI that reads your request, decides what to do, and (in agent mode) takes actions on your behalf.

The agent always has context: it knows which app is in front, can read your selected text, and (when you ask) can see your screen, your browser tabs, and your working directory.

Modes

The same launcher behaves differently depending on the mode:

  • Chat (default): a direct conversation with a model. No tools; fast answers.
  • Agent: the model can use tools to act: browse, search, run code, edit files, control apps. Sensitive actions require your approval.
  • Vision: the model can see images and screenshots you attach.
  • Search: the agent runs live web searches and refines its queries.
  • Imagine: generates images from a prompt.
  • Think: asks reasoning-capable models to show extended thinking.

You switch modes with a tag (/agent, /search, /imagine, …) or a keyboard shortcut. See Launcher Modes for the full list.

Tools

In agent mode, the agent calls tools to interact with the world: web browsing, the file system, shell, Python, Chrome control, Calendar, Mail, iMessage, and more. You enable tools in Settings → Tools and Plugins, and you can extend the set with MCP servers. The agent shows each tool call before running it so you stay in control. See Tools & Integrations.

Models and providers

A provider is where models come from: Anthropic, OpenAI, or the built-in Agentastic cloud provider. A model is a specific brain (Claude Opus, GPT-5, Gemini, …). You can:

  • Pick a model per request with /model: (e.g. /model:opus).
  • Let Agentastic route automatically with agentastic:auto.
  • Set per-role defaults: different default models for chat, fast, agent, vision, computer use, image, dictation, and bundle comparisons.

See Models & Providers and Choosing Models.

Skills

A skill is a reusable package of instructions that teaches the agent how to handle a recurring workflow. Skills come from Skillopedia, a public catalog, and they can auto-activate based on rules: the app you're in, the website, or keywords in your prompt. See Skills.

Memory and channels

Agentastic keeps memory so it remembers facts and preferences across conversations. Memory is organized into channels (like #work or #personal) that you and your agents can subscribe to. See Memory & Channels.

Agents (the plural)

Beyond the default assistant, you can create specialized agents, each with its own name, personality, default model, pinned skills, memory, and subscribed channels. You summon one with @agent-id in the launcher, or bind it to a schedule so it runs on its own. See Agents.

The Gateway

The Gateway connects your agent to messaging platforms (Slack, Discord, WhatsApp, Telegram, iMessage, and email) so you can talk to it from your phone or a team channel. See Gateway.

The background service

A lightweight daemon runs in the background. It hosts the agent, executes tools, runs scheduled tasks even when the launcher is closed, and powers the Gateway. The menu bar icon and the agentastic CLI both talk to this service. See CLI Commands.

How it fits together

You ──hotkey──▶ Launcher ──▶ Agent ──▶ Model (via Provider)
                               │
                               ├─▶ Tools (web, files, shell, MCP, …)
                               ├─▶ Skills (auto-activated by rules)
                               └─▶ Memory & Channels

Background service ──▶ Schedules · Gateway (Slack, iMessage, …)

Privacy

You decide which providers to use, the agent asks before taking sensitive actions, and you control which tools and folders it can touch. Local models via MLX let you keep inference on-device entirely.

What's next