When Should You Keep a Claude Code Host?

Try a Claude Code cloud handoff for branch-based jobs. Compare it with persistent hosting on paired recovery drills, and value the time saved only after both paths produce the required files, context, and test result.

Make the handoff prove what it recovered

Try a Claude Code cloud handoff with a disposable branch, a pinned revision, and a small deterministic test. Include a harmless untracked fixture and a synthetic instruction remembered only in the original conversation. From the receiving environment, ask for the fixture checksum and the instruction, then run the test and record all three results. Mark each item required or optional before running the drill. A branch-based workflow can pass with committed fixtures and a written handoff note. A task that requires untracked artifacts needs a recovery path for those files as well. Anthropic’s setup excerpt specifies Git 2.31 or later for uploading a local repository to a cloud session from macOS, Linux, or WSL. Test that upload and first successful run from the least configured workstation you expect operators to use. Record setup time separately from the agent’s work time.

Advanced setup - Claude Code Docs

Test a replacement host before valuing persistence

For a self-hosted Claude Agent SDK deployment, run a second drill on an empty replacement host using the same working-directory path. Restore the repository and the required workspace artifacts, then resume through your SessionStore adapter. Anthropic describes the adapter as a transcript mirror. Check that the resumed session recalls a synthetic instruction, the worktree has the expected checksum, and the test exits successfully. A transcript that resumes alongside a missing fixture identifies a gap in workspace restoration. Record elapsed time and hands-on recovery time separately. In an illustrative workload with 20 recoveries a month, saving nine hands-on minutes per recovery saves three operator-hours. Multiply by your operator cost and compare that figure with the monthly difference in infrastructure and model spend. Anthropic’s published plan page groups Claude Code and other plan usage, so measure incremental model spending separately from the host bill.

Persist sessions to external storage - Claude Code Docs

Test the cloud handoff from a fresh session

Start a Claude Code Cloud pilot with a disposable repository and one task that ends in a reviewable branch. If your workflow uploads a local repository from macOS, Linux, or WSL, run git --version first. Anthropic's setup documentation requires Git 2.31 or later for that upload path. For a GitHub-connected pilot, check that the intended repository and branch are available before assigning work, then verify the resulting branch independently in GitHub.

Close the client during the run and check whether you can return to see the outcome. Then start a fresh session and ask it to reproduce the result using the branch and a short handoff note. If reopening works but the fresh session needs a file you never exported, you have found a state dependency that client reconnection alone cannot resolve.

Advanced setup - Claude Code Docs

Price the login and the host separately

Anthropic's pricing excerpt lists Claude Pro at $20 when billed monthly or $200 billed up front for an annual subscription. It also says Claude Code and Claude activity across web, desktop, and mobile draw from the same usage pool. During the pilot, check which account the coding session uses and record any usage-limit interruptions alongside the number of completed tasks. Confirm plan eligibility, usage credits, and any cloud-specific charges on the official pricing page and in the account that will run the jobs.

For an illustrative comparison, assume one $20 monthly subscription and a separate host quoted at $12 per month. If the same subscription covers both paths, the persistent option starts at $32 per month under those assumptions. Verify the host's billing unit, disk and backup charges, and whether its sign-in method uses that subscription before relying on the total.

Plans & pricing | Claude by Anthropic

Prove that the workspace can be rebuilt

For a self-hosted Claude Agent SDK deployment, the SDK supervises a Claude CLI subprocess with its own shell, working directory, and session files. Inventory the paths your task depends on before picking storage. Keep source changes in version control, keep necessary generated artifacts in recoverable storage, and document how to rebuild disposable dependencies. Decide separately whether any running process must stay alive between turns.

For an illustrative case, assume a Monday job finishes a parser change and leaves a 600 MB generated test fixture on disk for Tuesday. Put the commit on a branch and store the fixture with a checksum. On a replacement host, check out the commit, retrieve the fixture, and rerun one test. A matching checksum and passing test support a rebuildable setup. If Tuesday's work depends on a still-running process, test a persistent host and define what happens when that process exits.

Hosting the Agent SDK - Claude Code Docs

Sources