Assistants
An assistant is a reusable persona: it bundles display info, a default agent backend, an optional model preference, a system prompt, and a skill selection into one entry, so you can pick it in one click when starting a session instead of rewriting prompts every time.
The entry point is /assistants (the old /settings/assistants redirects to /assistants?tab=assistants). The Skills tab on the same page lives at /assistants?tab=skills.

Steps
-
Browse the assistant library. Open
/assistants. The list merges three sources: Builtin (shipped with the app), Custom (created by you), and Extension (provided by installed extensions). Every entry can be enabled/disabled and reordered, and the list also tracks recent use so you can grab one quickly in a session. -
Create an assistant. Click new, then fill in a name, description, and avatar; choose a default agent backend (e.g. the built-in
nomi, orclaude/codex/gemini); set a model preference if you want; write the system prompt (the assistant’s instructions); and check the skills to attach when a session starts. Save, and it’s ready to pick in a session.
-
Edit an assistant. Custom assistants are fully editable — name, description, avatar, prompt, and skills — and can be deleted. Builtin assistants are read-only in content, but their enabled state, sort order, recent use, and a default-backend override are stored separately; Extension assistants are read-only here, with their lifecycle managed by the extension.
-
Organize with tags. Create tags in tag management and apply them to assistants, then filter by tag at the top of the list — once you have many assistants, this is the fastest way to find one.

-
Pick one in a session. When you start a session, or switch agents in the session header, choose an assistant; its prompt and enabled skills take effect automatically at session start. The assistant’s skills merge with the auto-injected builtin skills — no need to copy skill directories by hand in normal use.
Notes and limits
- Three sources, different editability. Custom assistants are editable and deletable; Builtin assistants are read-only in content (only enabled/order/default-backend override are mutable); Extension assistants are read-only.
- Default backend and providers. If you don’t set a default backend when creating an assistant, you need at least one configured provider; the built-in
nomiis used by default when it can be inferred. - CLI agents need their own install. Choosing
claude/codex/geminidoes not install those tools — they still require the matching CLI on the host beforehand. - Skills are per-companion; memory is shared. Each companion has its own skill library, while memory runs through one capture/learning pipeline shared by all companions. Per-companion private memory is coming soon.
- Skills and MCP are managed separately. The Skills tab is at
/assistants?tab=skills, while MCP servers moved out to/mcp; both extend capability but are managed apart today.
FAQ
- How do assistants relate to sessions? An assistant is a template; a session is an instance. Picking one makes its prompt, default backend, and skills the session’s starting point; later tweaks inside the session don’t rewrite the assistant itself.
- Can I edit a builtin assistant? Not its content, but you can enable/disable it, reorder it, and override its default backend. For full customization, create a custom assistant instead.
- What happens when I delete a custom assistant? Its associated rule, skill, and avatar files are cleaned up along with it.
Related
- MCP and skills — manage skills and MCP servers to extend what an assistant can do.
- The session workspace — pick an assistant and get work done in a session.
- Model failover queue — a safety net for an assistant’s model preference.
Full docs → GitHub