A cloud for agents, with the work kept together.

Use cloud resources around a concrete deliverable

An agent cloud becomes useful when it helps finish work: a source-based brief, a checked spreadsheet, a private App, or a repeatable specialist. Start with one deliverable and keep its sources and review criteria in a Project. Add compute or a cloud browser when the task needs them rather than making every request a machine-management exercise.

For example, a consultant can keep approved notes in a Project, ask ChatGPT to prepare a saved report, and ask Claude to review the conclusions. A technical operator can connect Hermes or OpenClaw to the same authorized Workspace tools. Each client still needs its own working connection; the Project does not synchronize private conversations between providers.

Choose the assistant and connect its tools

Grow from files to Apps and reusable Agents

Use saved-work tools for Projects, files, task records, and handoffs. Discover native tools when the job needs Office files, Agent configuration, compute, a hosted browser, or App work. Tool discovery returns the current schemas and required permissions; a catalog entry does not prove that every account, provider, or client can run it.

Create a custom Agent when the method should be reused. Define accepted inputs, selected tools, exception behavior, and a review checklist, then test it. Build a private App when people need an interactive tracker or dashboard. A draft, a ready preview, a published App, and a successful Agent run are distinct results; verify the one the user requested.

Keep access and usage understandable

The MCP connection uses browser OAuth. Read is the starting point; Edit permits saved-work changes, while Configure and run enables execution and configuration within the user’s role. Choosing a Project does not turn a broader account-role grant into a Project-only credential. Plan checkout is a separate permission, not a prerequisite for a connection test.

Your external assistant’s model usage remains with its provider. Clanker-hosted inference, compute, builds, and other enabled operations follow their applicable usage rules. A client subscription does not make all Workspace work free. Inspect account usage and stop environments or recurring workers you no longer need.

Sources: Clanker: connect an external agent and review permissions · Clanker: current usage and plans

Choose the cloud interface that matches the job

Use Workspace MCP when an external assistant should work with your Projects and authorized tools. Use a supported hosted Agent environment when you want to configure agent execution in Clanker. Use the standalone Sandbox API when your own application needs a programmatic runtime under that separate API contract.

Use a connected computer only when the task needs that particular machine or local application. A connection to Workspace does not expose your local desktop automatically. Recurring Loops and Schedules also require their own configuration. Keep these choices explicit so saved context, active processes, permissions, and costs are not confused.

Try it with a small example.