Use cases
Real work, on the machine that is ready for it.
Choose the narrowest workflow that gives an Agent the context it needs. Expand access only when the work requires it.
Continue local development
Let an agent inspect a repository, apply focused edits and run existing tests without rebuilding the environment elsewhere.
- Typical access
file.read · file.patch · test.run- Expected outcome
- The existing repository, dependencies and toolchain stay on the selected machine.
Connect existing MCP tools
Register local stdio or Streamable HTTP MCP servers without moving Provider credentials to the control plane.
- Typical access
GitHub MCP · database MCP- Expected outcome
- The Agent receives only requested results rather than a copy of the whole environment.
Move work across interfaces
Start from an MCP client, terminal, Skill or SDK and route the same governed work to the intended device.
- Typical access
MCP · CLI · Skill · SDK- Expected outcome
- Every interface uses the same authorization, error semantics and audit trail.
Automate with boundaries
Use explicit templates, writable folders and approval rules for repeatable work instead of sharing unrestricted access.
- Typical access
template · approval · receipt- Expected outcome
- Sensitive actions remain reviewable and retries have explicit idempotency semantics.
Agent Device Cloud is not a public compute marketplace, a general RPA product or a hosted shell.