Security, permissions & provenance

Signed-in chats, files, and account data are stored against your account so they can sync across the web, Android, Windows, and CLI. Database access rules restrict each chat to its owner, except for content you explicitly share or invite someone to collaborate on.

Browser caches and local drafts are convenience copies, not the source of truth. Clear drafts from the composer or delete the conversation from Nexara when you no longer need it.

Claw uses three operating modes: Default asks before changing or destructive actions, Accept edits permits ordinary project edits while retaining gates for higher-risk actions, and Bypass permissions removes those interactive gates. Choose the least permissive mode that fits the task.

External services, outside-project files, terminal commands, publishing, and financial actions are permission-sensitive. Each action records a local audit event with its type, time, result summary, and whether approval was required; sensitive payloads are reduced to metadata rather than copied into the log. The local runtime retains the latest 500 events.

Agent work also has independent Compute guardrails in Settings → Privacy & Data: a daily cap for scheduled/delegated work and a per-run output cap. These are enforced server-side; reaching a cap skips the run and records the reason. They do not increase your plan limit, and ordinary chat remains governed by the plan's daily and weekly limits.

A label such as openai/GPT-5.6Luna is a Nexara catalog identifier: the first part describes the routing namespace and the second is the model label Nexara displays. It is not a claim that OpenAI or another upstream provider sells a retail product under that exact combined name. Model cards and replies should be read with that distinction in mind.

Stop an active Claw run, revoke the connected service or local permission, and review the audit timeline before continuing. For account, deletion, or security concerns, use the support channel.