An assistant is a chat partner: you send a message, it streams back a reply, one turn at a time. Every assistant message is its own "run" — it starts, it finishes (or fails), and it's checked against your plan's runs-per-day quota.
An agent works toward a goal across multiple steps, not a single reply. You give it a goal; it runs — in the foreground (you watch it stream, like a chat) or in the background (you check back later) — calling tools and, if it needs to, pausing to ask you a clarifying question or wait for your approval before a sensitive action (like sending an email). An agent can also delegate part of its work to another agent you've attached as a sub-agent.
Use an assistant when the interaction is naturally a conversation. Use an agent when the work is a task with steps — something you'd otherwise have to babysit through several back-and-forth messages yourself.
A workflow runs fixed stages the same way every time: you pick one of five patterns (chain, routing, parallel, orchestrator-worker, evaluator-optimizer) and fill in its prompts and models, instead of letting an agent decide the steps. Use one when the stages really are fixed and you want a predictable result. See Workflows.
A workspace (what the product calls an "workspace") is where assistants, agents, connections, and knowledge sources actually live — it's the unit you create, name, and invite teammates into. A personal workspace is just for you; convert it to a team workspace when you want to add members.
A billing account is what actually pays. Your personal workspace and any team workspace you create both belong to the same billing account, and every plan limit (assistants, agents, workspaces, members, runs/day) is checked against the sum across all of them, not per workspace — otherwise you could dodge a limit just by creating another workspace.
Most people only ever see one workspace and never think about this distinction. It starts to matter once you have more than one.
Plans (Free, Solo, Pro, Team, Enterprise — see Pricing) bundle a fixed number of included credits plus usage limits (how many assistants/agents/runs-per-day/workspaces/members you get). Credits are consumed by platform-key model usage, priced per model at the provider's own cost plus a markup, never below a small minimum fee per run.
Credits top up two ways: automatically, whenever your plan renews or you change plans, and manually, with a one-time purchase independent of your plan. Both land in the same balance — there's no separate "bonus credits" bucket.
See Billing & credits for the full mechanics.
BYOK ("bring your own key") means you've connected your own API key for a provider (OpenRouter, DeepSeek, OpenAI, Anthropic or Google) to a workspace. Any assistant/agent using that provider then routes through your key instead of the platform's — you're billed by the provider directly, and credits are never touched for it.
Platform keys mean the platform's own provider key is used, and your credit balance is charged per run. This only works for models the platform has explicitly priced (the curated model list) — an arbitrary model id isn't billable on the platform-key path even if the provider supports it.
You don't pick one globally: BYOK is set per workspace, per provider. A workspace with its own OpenAI key but no Anthropic key uses BYOK for OpenAI models and the platform key (if available) for Anthropic ones. See BYOK.