Skip to content
Browse documentation

Permissions & safety

Control project access and protect data

Make least-privilege decisions for local tools, providers, and exports.

Vopit separates local project permissions from agent capabilities and from infrastructure grants. The permission categories are READ, WRITE, EXECUTE, NETWORK, GIT (inspection), GIT_WRITE (branches/commits), GIT_REMOTE (synchronization), and GIT_HISTORY (merge/rebase). Project rules override global rules.

Respond to a permission request

  1. Read the requested operation and its project scope. Treat commands and writes as real operations.
  2. Choose Deny, Allow once, Allow for project, or Always allow. The last two apply to the whole permission category, not just the displayed command; use them only when the scope and agent are trusted.
  3. Review or reset a project override in Settings → Permissions when the task ends or the project changes.
  4. For a native Codex stage, Vopit grants the permission categories required by the stage; additional sandbox requests return to Vopit authorization.

Answer an agent’s question

During its task, an agent can ask you a question when it needs your decision or information: one option, several options, or a free answer, with the context it found. Its stage waits for your answer without using its time limit. The question opens in Vopit, appears in the terminal (vopit questions, vopit answer <id> <answer>) and on Vopit Remote.

  • Pick an option or write your own answer when the question allows it, then choose Answer. The agent reads it and continues.
  • Skip lets the agent continue with its best judgment; it states the assumption it made in its result.
  • Settings → Permissions → Wait for an answer sets how long agents wait before deciding on their own: until you answer (the default), or from 5 minutes to 1 day.
  • Every agent may ask by default. To stop one from asking, turn off Ask you questions in its editor, under Vopit actions.

Default rules and once-only prompts

  • Settings → Permissions offers Ask every time, Allow for project, Always allow and Never allow for each category. By default READ is Allow for project and every other category is Ask every time. Project rules take priority over these global rules.
  • Infrastructure prompts offer only Deny or Allow once: approval covers that execution only. Future approvals are managed per operation in Connections.
  • Crank action approvals use a once-only NETWORK prompt. Each call to an MCP server granted to a Claude Code agent asks for NETWORK permission.
  • Setting READ to Never allow also prevents AI generation features from reading project files, such as Read this project to fit the team to it and Use the open project's files.

Data and secret handling

  • Vopit project tools protect `.vopit` along with credentials and other internal project data. Local Vopit shell commands are not container-isolated. Native Codex sandbox behavior and reads are separate and do not inherit Vopit content filters.
  • Provider-bound task context can be transmitted to the selected provider. Choose the provider according to your organization’s data policy.
  • Request/event and log exports may contain project data. Keep them in approved storage. No import or restore interface for those exports is verified.
  • Use a dedicated SSH account with limited privileges for deployment. An infrastructure inspection label and a remote root directory do not sandbox the command.

All documentation