Everyday operating habits
Stay in control
of access and cost.
Know which machine is doing the work, which accounts it can reach, and which usage it consumes.
New to modes? Read the Permission Modes comparison and dedicated Auto, Plan, and Accept Edits guides.
Understand permission modes
Use the six-mode comparison to choose deliberately. Start with Plan for investigation or Manual for step-by-step approval. Cloud sessions expose Auto, Accept Edits, and Plan. Check the current indicator before each task.
Grant the access the task needs
- Start Claude in the specific project folder, not your entire home directory.
- Use
/permissionsto inspect existing allow, ask, and deny rules. - Use separate test credentials and development databases.
- Restrict GitHub apps and tokens to the intended repositories.
- Inspect hooks, plugins, MCP servers, and project setup scripts before trusting them.
- Use a separate user, container, or VM when stronger isolation is needed. Containers still need careful mounts and privileges.
Permission rules and sandboxing address different layers. A deny rule on one file-reading tool is not a complete defense if another allowed tool or external integration can reach the same data. Treat repository instructions and fetched content as untrusted inputs until reviewed.
Where your information goes
Local execution can still send relevant code, conversation content, and tool results to the model provider. Remote Control additionally synchronizes session content through Anthropic. Third-party integrations have their own data access and retention behavior.
Consumer and commercial account terms can differ. Review your actual account’s privacy settings and the current data-usage documentation. Do not assume that using a private repository, a home server, or a paid account makes every request offline or zero-retention.
Use secret stores or approved authentication flows. Remove keys, access tokens, personal information, and production records from screenshots and logs before sharing. If a secret leaks, rotate it; hiding the chat or deleting one file is not enough.
Cloud session sharing: on Pro and Max, Public means any signed-in claude.ai user can view the session. The repository-access check is off by default; inspect the sharing controls before exposing code or conversation history. Auto-fix can reply under your GitHub account and those replies can trigger comment-based automation.
Understand the billing paths
| Usage path | What to check |
|---|---|
| Claude subscription | Included usage, model availability, limits, and any enabled additional usage. |
| Console API key | API billing, workspace budget controls, token usage, and whether a key overrides the intended login. |
| Cloud provider | Provider-specific authentication, billing, supported models, and regional settings. |
| Automation and agent teams | Each execution can consume usage. Parallel work can multiply requests. |
| Home hosting | Power, backups, hardware, domain fees, and any external services used by the app. |
Use /status to confirm the account and /usage to inspect usage. Exact prices and limits change, so consult current Claude pricing and your billing dashboard rather than relying on a fixed dollar estimate here.
Choose a suitable model in /model. More reasoning or a faster service tier can change consumption and billing. Give focused tasks, inspect /context, and use /clear for unrelated work in local sessions. Cloud sessions do not support /clear; start a new session from the sidebar instead. Use /compact when you need continuity in a long conversation. Compaction itself processes the conversation.
Keep recovery outside the conversation
Commit reviewed changes in Git and back up files you cannot replace. Claude checkpoints can restore eligible file edits and conversation state, but do not undo every shell command, remote action, database migration, deployment, or external purchase. Test database and server backups separately.
Before a change with lasting effects, decide how to restore the prior state. A successful review reduces uncertainty; it never eliminates the need for backups and independent checks.
- Anthropic. (n.d.-e). Checkpointing. Retrieved September 24, 2026, from https://code.claude.com/docs/en/checkpointing
- Anthropic. (n.d.-l). Configure permissions. Retrieved September 24, 2026, from https://code.claude.com/docs/en/permissions
- Anthropic. (n.d.-n). Continue local sessions from any device with Remote Control. Retrieved September 24, 2026, from https://code.claude.com/docs/en/remote-control
- Anthropic. (n.d.-p). Data usage. Retrieved September 24, 2026, from https://code.claude.com/docs/en/data-usage
- Anthropic. (n.d.-ab). Manage costs effectively. Retrieved September 24, 2026, from https://code.claude.com/docs/en/costs
- Anthropic. (n.d.-ao). Use Claude Code in the cloud. Retrieved September 24, 2026, from https://code.claude.com/docs/en/claude-code-on-the-web
Put security into practice
Follow the beginner security path, then learn server-side application controls and advanced release reviews. Each lesson includes an implementation prompt and a way to check the result.
Was this chapter useful?
Report an errorVotes are shared only if you enable optional measurement in Privacy & Preferences.