One plan, the right model for each part

Giving an agent a whole job is only useful if you can see the plan, limit what it can touch and review what comes back. That’s how team runs in Locus are designed.

A bounded job graph

A dispatcher breaks your goal into a bounded job graph, for example: inspect the current implementation, implement and verify the backend, then build the UI against the finished backend contract. Each specialist gets its own model and an access ceiling, so you can run one role on a local model and another on an account you’ve connected.

One approval, with rules

One approval covers the listed jobs and their read-only helpers. Agents that change files can’t delegate, and coding jobs run one at a time in the shared checkout. Tool actions follow your shared approval policy.

• Limits on jobs, rounds, concurrent calls, total calls, metered tokens and estimated cost
• Optional managed Git worktree, so you review the diff before it reaches your project
• Conflicts checked before applying, or the result lands on a branch

Local by default

Your Mac hosts the workspace, so files, sessions and prompt history stay with you. Hosted routing is authorized before a dispatcher sends task context to a provider. Locus needs no SparkTales account and sends no analytics or telemetry to SparkTales.