Sanbox By Sanlabs

Guides

Agents for teams

Available on request

See and cap what each AI agent spends

By the Sanbox team updated

Ask ChatGPT Ask Claude

Sanbox shows what each Computer and each team spends, and lets you set a cap for each one, so a busy agent or a busy team can't run up a surprise bill. Spend caps are enabled per organisation on request. Today, hosting is agreed per organisation, and model spend is controlled with your model provider's own limits.

What do you get?

  • Spend per Computer. See what each agent costs, not only the organisation's total.
  • Spend per team. Group Computers by team and see each team's share.
  • Caps you set. Give a Computer or a team a limit for the month, and Sanbox stops new work when it reaches the cap.
  • Spend you can follow. Owners see spend build up during the month, while there's still time to act.

This is for organisations that run many Computers: a Computer per employee, an agent per customer, or a fleet of agents doing background work.

Enable it for your organisation

Enable it for your organisation and tell us how your Computers are split into teams and what limits you need. We'll set it up with you.

What works today

Hosting agreed per organisation. Sanbox plans are agreed with each organisation. Talk to us about a plan for your team.

Model spend lives with your provider. Sanbox doesn't sell models. Your agents call the provider you connect, with your key, so the provider's own limits apply. For example, Anthropic's Claude Console lets you set a monthly spend limit for the organisation, and separate spend limits for each workspace except the default one. When a limit is reached, the API refuses further requests until the limit resets or you raise it.

One key per team. To keep teams apart today, give each team its own model key with its own limit at the provider or in a gateway you run, and connect it to Sanbox as its own custom OpenAI-compatible endpoint. Use that endpoint only in the team's templates, and the limit on that key becomes the team's limit. Adding a custom endpoint needs the owner or admin role; see How do I use my own model endpoint with an agent computer?

Stop what you don't use. A Computer runs until it is stopped; nothing puts it to sleep. Stop Computers nobody is using: a stopped Computer makes no model calls. Its disk is kept for when you resume it.

What to know

  • Caps on a model key stop the agent mid-task. When a provider limit is hit, the agent's next model call fails. Size limits so normal work fits.
  • A shared key means a shared limit. If several teams use one provider connection, one limit covers all of them.

Sources