Claude Code for teams

What the team is no longer doing

A local session stops when the laptop sleeps, the permission prompt is answered by whoever is sitting there, and the only record is that machine. A team handing Claude Code real tickets needs something else: the session continues without that laptop, the approval waits for the person who should decide, and the pull request shows up with the spec, the review notes, QA’s results, and the cost.

AOSplane does not replace Claude Code with a different agent. The harness is Claude Code. CLAUDE.md in the repository still applies. The credential is issued for the run and is not stored in the sandbox. Model calls use the Anthropic account, Bedrock, Claude Platform on AWS, or another compatible provider you connect. A Claude subscription sign-in is rolling out and is always personal.

Several agents, one gate

Define more than one agent when the ceilings should differ. A builder that iterates on tests needs a higher cap than an agent that only reads. Each definition has its own harness, model, instructions, and budget. You can put Claude Code on one and Codex on another. The review and the QA pass do not depend on which one built the change. Matching a task to an agent automatically is still being built. Today you choose.

Work arrives from the console, from a Linear issue, from a Jira Cloud issue, or from a project API key in one project and one environment. Slack is rolling out: mention @AOSplane or use /aos. A channel can carry its own project, agent, and rules, and it may not be on for every organization yet.

The hard cap

Shared AI-credit balances, the kind Linear’s workspace credits, Copilot’s pooled credits, and Rovo’s allowances use, let one session spend what the rest of the team had left. AOSplane checks the organization, the project, and the agent before the run. The tightest limit is the task’s cap. The task does not start if the reservation would cross it. A run already in flight is not cancelled when a later one is refused.

Claude Code has its own limits too: a session flag, a Console workspace limit, and spend limits on a Teams or Enterprise plan. Those are Anthropic’s. They are not replaced by the task cap. How to set each one. The agent guide is Claude Code in the cloud. Codex beside it is Claude Code vs Codex.

Who still merges

The team merges. Agents cannot push to the repositories. The draft opens after review and QA pass, and a person decides. Sensitive steps stop and wait in the console. The decision is on the task’s record, with the reservation and the pull request.

The Free plan is the self-serve way to see it: 2 projects, 2 tasks at a time per project, small machines, 7 days of history, and a $5 monthly usage credit. Starter is 5 projects and 8 concurrent runs. Team is 25 projects and 32. The plans. What the team is no longer re-checking by hand.