Assistants

An assistant is a configured chat partner, scoped to a workspace. See Assistants vs. agents if you haven't read that yet — this guide assumes you have.

Configuration

• Provider & model — which provider/model it runs on. Only curated, priced models are available unless the workspace has a BYOK connection for that provider (see BYOK).
• Instructions — the system prompt: who it is, how it should behave, what it shouldn't do.
• Core memory — a persistent block always injected into context. The assistant can append to it or replace part of it itself (via its own memory tools), so it can "remember" things across turns without you editing the instructions.
• Knowledge mode — whether, and how, it searches an attached knowledge base.
• Max steps / max output tokens / temperature — the usual generation controls, plus a cap on how many tool-call round-trips a single reply can take before it has to answer.

Chatting

Every message you send is one run — it streams back a reply, then finishes. Each run counts against your plan's runs-per-day limit (see Billing & credits).

Tools

Attach any custom tool, tool from an MCP server, or built-in capability. A tool marked as requiring approval pauses the reply mid-stream until you approve or deny the call, rather than running immediately.

Skills

Attach skills on the Edit tab to give the assistant instructions it loads only when a task needs them.

Memory

Beyond core memory, an assistant always has access to archival memory — it can save a note for later (remember) and search its own previously saved notes (archival_memory_search), independent of whatever's in the current conversation.