Superagent
idle
Privacy

What can I take off your plate?

I read your own numbers, keep what you teach me, and only act after you approve.

Library

Marketplace · soon

Files

—

Chats

—

Skills

0 skills

Written guidance the agent loads when it's relevant. Only names and descriptions sit in every prompt — the full text loads on demand, so guidance can be thorough without costing tokens on every message. Same format as Claude Code skills (SKILL.md + optional reference files), so skills are portable both ways.

Memory

0 notes

Your agent's identity — loaded in full at the start of every session. Keep it short and durable; topic knowledge belongs in a note.

Start from a template:

This is GOALS.md as the agent reads it, loaded into every session alongside the soul. Goals moved to their own page — — and this is the file behind them, for a hand edit. Keep each - Id: line: it links a goal to its plan.

Edit as markdown

Coworkers

…

Goals

…

Connections

Connect an account and the assistant reads your own numbers instead of asking you for them.

Add Note

A page in the agent's notebook. The summary is what it sees in the index — make it say whether the note is worth opening.

New Skill

What these settings mean

Model tiers

A tier is a name for a job, not a model. The server picks the model and always the same one, so a chat started on “Balanced” today runs what “Balanced” ran yesterday. Picking a model by name instead pins it exactly.

Either way the choice applies to your next new chat — a chat keeps the model it started on.

Tool limits

Models are certified or uncertified. Certified ones have been run through a live checklist — real tool use, obedience to the fence around untrusted text, cancellation.

An uncertified model is offered a read-only set of tools: it can read, search and browse, but no writing, no shell, no connectors. Not refused at the moment of use — simply never handed to it, because a tool that is described and then refused is an invitation a model keeps accepting.

As the owner of this deployment you are the exception: uncertified models get your full tool set, on your data, at your risk. Everyone else gets the read-only set until the model is promoted.

Costs and the ~

Every figure here is computed locally by the agent runtime. It is not a bill from anyone.

On Anthropic models it is arithmetic on published rates and lands close. On models reached through the OpenRouter gateway it is shown as ~ because the runtime prices them at a stand-in model’s rates — a measured run came out 3.8× higher than the real invoice. It errs high, so a spending cap stops you early rather than late. Reconcile real spend on your provider’s own dashboard.

Approvals

Actions that change something outside this conversation pause and ask. Nothing runs while a request waits, and silence is a refusal — an unanswered request is denied, never approved by timeout.

Approval is per action. There is no “always allow”.

Settings

Stored in your browser only. Sent as Authorization: Bearer on API calls.

Appearance
Anthropic account

Your brands

One login, several businesses. Each brand keeps its own soul, goals, notes and files, and none of those is shared. A connection is the exception: connect a service once and choose which brands may use it. Switching reloads this page.