Skip to main content

Set Up a Team

Open Settings → Agents & Teams.

1. Create agent profiles​

A profile defines:

  • name and role;
  • exact provider account and model;
  • role instructions and capability tags;
  • read-only, workspace-write, or Computer Control ceiling;
  • standard tool groups and explicit MCP allowlists; and
  • timeout, token, call, and optional estimated-cost limits.

Use Test Connection before adding a profile. Eligible routes include Ollama, ChatGPT-plan, OpenAI API, Claude API, Kimi, and compatible endpoints, plus Claude plan where the 3.0 preview capability and its managed helper are available. ChatGPT-managed teams keep the Locus team contract rather than Codex-native Solo behavior.

2. Define the team​

Choose one dispatcher, an optional fallback dispatcher, one write-capable Lead Writer, any additional ordered writers, and read-only specialists or reviewers. Configure maximum jobs, rounds, simultaneous calls, total calls, metered tokens, and estimated cost.

New teams use adaptive Locus execution with bounded delegation. The optional OpenAI Responses beta appears only for an eligible OpenAI API GPT-5.6 dispatcher, never a ChatGPT-managed account. If it is unavailable, Locus pauses and offers to run with the Locus engine; billing and execution never switch silently.

3. Allow hosted routing​

Review Settings → Agents & Teams → Automatic Hosted Routing. Consent names the provider accounts that a dispatcher may use without another routing prompt. It does not grant tools, raise an access ceiling, or bypass ordinary permissions.

Starter team​

A practical first team uses:

  • one local or inexpensive dispatcher;
  • one read-only researcher;
  • one primary writer;
  • one read-only reviewer.

Keep initial budgets small until the routing and evidence match your expectations.

Project checks and saved-agent chats​

Manage reusable checks in the same settings page opens the current project's approved checks and proposals. New proposals begin with Make reusable check on a saved user message; opening settings does not create one.

In 3.0, a saved-agent chat can select an explicit team or delegate bounded helper work. The chat retains its saved-agent owner; each stage remains bounded by its configured specialist profile and the chat’s permissions. Team stages use their chosen routes and budgets. A profile's default is Read only. Choose Workspace edits for a writer, and remember that the chat's permission mode can impose stricter limits.