Skip to content

Concepts

The words you meet in Handler, a few lines each, with the page to read next. Contributors: every rule behind them is in the domain model.

 model access ─────────────┐  an account (your subscription) or a provider (an API, a local model)
                           ▼
 project ──► agents ──► sessions ◄── you: the web UI, a terminal, an editor, your phone
 a folder    one goal     one conversation each, in a sandbox
             each, with
             its settings

Model access

Account. A subscription an agent program signs in with: Claude Pro or Max, a ChatGPT plan, a Mistral account. For your own use only. Settings → Model access.

Provider. Where model calls go when it is not a subscription: an API with its key (Anthropic, OpenAI, OpenRouter) or a model on your machine or network (llama.cpp, Ollama).

Harness. The agent program that runs: Claude Code, Codex, Mistral Vibe or pi. An account brings its own; on a provider you choose one.

Context. What a project is for (Work, Personal, a client). Accounts and providers say which contexts they are made for, so a work account never runs on a personal project.

Projects, agents, sessions

Project. A code folder and the agents you work on it with. The web UI's home page lists them.

Agent. Who works on a project, for one goal ("Dev", "Review"): its models, what it may do (network, permissions, sandbox) and what it gets (addons, skills, knowledge folders). Another model is not another agent: pick it when you start a session.

Agent outside projects. One made by handler run, handler acp or handler terminal in a folder that is no project's: the same command in the same folder gives the same agent.

Session. One conversation with an agent. Handler holds it, so it keeps running when you close the tab, and the web UI, a terminal and your phone all see the same one. Stopped after 30 minutes without anyone, it resumes when you come back.

Agent with an interface, and its runs. An agent you start from a form instead of a chat: it works unattended and gives back a result. Handler ships two: the skill reviewer and the skill creator.

Safety

Sandbox. Where an agent runs: a container that sees its folder and what you give it, nothing else of your computer. What it protects and what it does not: security.

Safety preset. Locked down (no network, asks first), Guarded (only its allowed hosts, never asks: the default) or Open. A starting point copied into the agent, then yours to change.

Network. No network, Filtered (only the allowed hosts; the session's network panel shows what was blocked and allows it in one click) or Open.

Permissions. Whether the agent asks before acting: Ask first, Auto edits, Auto-approve.

Work folders. The folders agents may work in at all (Settings → Folders, ~/code by default).

Secret. A token or a key. Agents never hold it: Handler adds it to their requests, towards the hosts it is for only.

What an agent can get

Addon and its configurations. An addon teaches agents one kind of system (a Shopify store, a browser): programs, skills, the hosts they need. A configuration is one set of its settings and secrets (one store, one token), and what you give an agent. Addons, a browser.

Skill. Instructions an agent reads when they apply: the Skills page, given by a project to all its agents or picked by one.

Knowledge folder. Your notes, given read-only to the agents that need them. Knowledge folders.

Toolbox. The tools every agent works with (compilers, runtimes, git…). Add one in Settings → Images → toolbox.

Around your agents

Automation. When this happens, run that: when a session stops or at a time, your script, one of your agents or one of your apps. Automations.

Inbox. What Handler tells you: your agents' results, your apps' notifications, failed runs.

Mini app. A small web app an agent writes for you and you deploy, used from your desktop or your phone. Mini apps.

Remote device and person. A phone paired with Handler through your gateway; it belongs to you (sessions and apps) or to a family member (only the apps shared with them). Your phone.

Words that are easy to mix up

Words Difference
agent / account / provider / harness The agent is who works, for one goal; it runs a harness (the program) on an account (your subscription) or a provider (an API, a local model).
addon / configuration / binding An addon is the package; a configuration is one set of its settings and secrets, what you give an agent; in files and on the command line it is called a binding.
agent / agent outside projects Both are an agent; one outside projects is made by the command line in a folder that is no project's.
session / run A session is a conversation; a run is a session of an agent with an interface, started from a form, unattended, with a result.
run / automation run A run is one of an agent with an interface; an automation's run is what the automation ran once (its script, an agent's run, an app's route).
session / harness terminal A session is held by Handler and shown everywhere; handler terminal is the harness's own terminal UI in the agent's sandbox, for as long as the terminal is open.

Where each of these is kept: where things are kept.