Tabs
Agent defaults
Agent and model defaults live on the project, not the account — there is no single account-wide agent default. Set them per project:- Agent, model, mode, and automation under Settings → Projects → Task Defaults. See task defaults for the full list. New tasks inherit the task defaults unless you override them in the composer. VM size is always selected at task creation; projects and accounts do not carry a hardware default.
Connections
The Connections tab contains integrations and Agent connections. The latter is one list: “Every plan, key and login you hold, each under its own name, and what runs on it.” See Agent connections for the options and default connection rules.Named rows and actions
Each row shows its name, option and Works with harness. Tags say what runs on it:- Default for Claude Code tasks, Default for Codex tasks, or Default for OpenCode tasks.
- Your agent, on the connection your personal agent runs on.
- <profile name> profile · <org>, for profiles pinned to it.
Add a connection
Choose Add to open Add a connection:- Pick an option tile. Each says which harness it works with; the same option can be added more than once.
- Enter Name, then complete Sign in for that option. A Claude plan asks for a Setup token; a ChatGPT plan uses Connect with ChatGPT and the device flow; API keys use Save key. An OpenCode auth file uses the displayed CLI command and Check status: “Run opencode auth login first, then the command above.”
--name '<name>'. The Claude plan hint is “Authorizes in your browser and stores the token.” The ChatGPT plan hint is “Opens the device login on your machine.”
Aether inference
The Aether inference row carries your organization’s balance and Add funds. Its Actions for Aether inference menu can make it the default for Claude Code or Codex tasks. No credential needs connecting. See inference balance for top-ups and the empty-balance refusal.Your agent
On Your agent, Runs on chooses Aether inference or a held connection by name, independently of your task defaults. The choices include the Claude plans, Anthropic API keys and OpenAI API keys your personal agent can run on. Unavailable options point to Connections. Choose its model alongside this setting; changing the connection or model restarts a live agent host.Agent profiles
The profile form has Name, Agent (“Select an agent”), Model (“Select a model”), Thinking level (“Select a thinking level”), and Runs on (“Select what it runs on”). Finish with Create profile or Save changes. Runs on offers Creator’s default (pin nobody), Aether inference where it serves the harness, and your connections by name. If a profile already pins a connection you do not hold, the field also shows that current pin with its owner, so saving other fields can keep it. PR reviews, automations and integration routes resolve through their profile: a pin chooses the connection; with nobody pinned, the task creator’s default applies.Task settings and turn records
A task’s connection is a setting like its model: chosen when the task starts, and changeable between turns. Runs on sits beside the model in the prompt bar and offers Aether inference or <name> · <option>, for example Work Max · Claude plan. The task’s setting says “applies to the next turn”; changing it does not change a turn already running. The new-task composer also has Runs on beside the model. Its reason says “your default for <Harness> tasks”, “nothing configured, so Aether inference by rule”, or “your choice for this task”. A task started through a profile that pins a connection runs on that pin, described as “the <profile> profile runs on it”; a choice in Runs on overrides it. A refused choice shows the server’s sentence under the select with a Connections link; an empty balance links Add funds. Each turn says Ran on: <name> · <option> or Ran on: Aether inference. A disconnected connection adds , no longer connected. If a turn’s segments differ, the line names them in order, A, then B. These are that turn’s records, not the task’s current setting.Platform API keys for calling the Aether API are managed under the separate API Keys tab, not Connections. GitHub, Linear, Slack and Sentry remain in the integrations section; see integrations.