Which fleet sandbox migrations can you approve?

Require an observed allow and deny check for each agent group’s destinations before moving the group to Docker cloud sandboxes. Use NanoClaw’s group layout to decide which agents need separate checks.

Inventory access by agent group

NanoClaw documents a workspace, memory, container, and selected mounts per agent group. Use the group as the unit of rollout. For each group, ask its owner for the destination pairs needed to finish a real job and for at least one destination the group should be unable to reach. Treat groups that use the same provider but handle different material as separate rows in the rollout plan. For example, a synthetic invoice intake job might require a test document service while a release notes job needs a test repository service. Using the same model for both tells you little about whether they can share network access. Preserve separate baselines when either group’s allowed destinations differ, and repeat the checks whenever you move an agent between groups.

GitHub - nanocoai/nanoclaw: Your own always-on AI assistant: self-hosted, open source, lightweight. Each agent runs in its own container on your machine, on the harness you choose (Claude Code, Codex, Cursor, OpenCode). Talk to it from Slack, WhatsApp, Telegram, Discord and more. · GitHub

Verify destination rules on the actual cloud sandbox

Docker documents separate account and sandbox cloud policies, with deny taking precedence when rules overlap. Local rules stay on the host during cloud creation. For each recorded destination, verify the connection from the cloud sandbox using an approved test endpoint and match the result to a policy decision in the logs. A timeout alone could mean the endpoint is down, so classify it as unknown until you have a policy decision or can repeat the check against a healthy endpoint. For example, an expected allow that resolves to deny can point to an account rule even when the sandbox policy view lists an allow. Record host and port explicitly. Use a separate application check for HTTP method or path restrictions because Docker’s cloud rules match destinations.

Manage cloud network policy | Docker Docs

Map permissions before assigning instances

Inventory each job's input, writable data, credential scope, schedule, destination, and owner. Separate jobs when a credential or file needed by one would be inappropriate for the other. NanoClaw provides an example of how to make that choice explicit. Its documentation describes separate agents per channel or an agent shared across channels, and per-group workspaces with their own memory, container, and allowed mounts.

In an illustrative fleet, two support channels serve Client A, a separate channel serves Client B, and a nightly review job reads an internal repository. Consider sharing an agent across A's channels if their permissions and retention rules match. Give B and the repository job separate execution boundaries. Test with synthetic files and narrowly scoped credentials. Each job should reach its assigned resources and fail to read a fixture assigned to another group.

GitHub - nanocoai/nanoclaw: Your own always-on AI assistant: self-hosted, open source, lightweight. Each agent runs in its own container on your machine, on the harness you choose (Claude Code, Codex, Cursor, OpenCode). Talk to it from Slack, WhatsApp, Telegram, Discord and more. · GitHub

Give each fleet cost an owner

Claude Code's cost documentation distinguishes API token consumption from subscription plan pricing. During a fleet pilot, log which job incurred model usage and keep hosting, model usage, and operator time as separate budget lines. Check the provider's pricing page or your agreement for the model or subscription you actually use. Check the host quote for the instance size, billing period, idle behavior, and storage charges.

For an illustrative month, assume three execution boundaries at $8 each, $35 in combined model usage, and one hour of maintenance at $60. The total is $119. If spend spikes, inspect it by job and attempt count before adding capacity. If two scheduled jobs repeatedly process the same input, fix scheduling or add duplicate protection, then run the fixture again and verify one accepted result.

Manage costs effectively - Claude Code Docs

Sources