Permissions
MCP and safety · AI agent permissions, permission modes, least privilege
AI agent permissions define what the agent may do in your environment and connected systems: only read or write as well, and which commands to run. The principle of least privilege applies: the agent gets exactly the rights the task needs, raised step by step.
Permissions in agents
- Steps · first read only, then writes to a test environment; writes to a live system go only through a person.
- Claude Code: rules · settings say what to allow, what to ask about and what to deny. A deny rule beats any mode.
- Claude Code: modes · by default it asks before edits and commands; acceptEdits accepts edits itself; plan only reads and plans; auto decides what is safe; bypassPermissions, the same as the
--dangerously-skip-permissionsflag, asks nothing and suits only a sandbox. - Codex: modes · read only; workspace mode, with edits and commands inside the project folder without asking and everything else with approval; full access, with no sandbox and no approvals.
Source: Claude Code permissions docs
Where it appears · 12
- AI agent = LLM + harness · Harness
- MCP server and AI agent for 1C · Harness
- Analytical reports with an AI agent · Harness
- Claude in Chrome · Harness
- The agent’s data plane and control plane · Harness
More 7
- Incident investigation with an AI agent · Harness
- QA work after support with an AI agent · Harness
- MCP for an AI agent · Harness
- Harness engineering in practice · Harness
- A reconcile loop for specs · Specs and context
- Specs by aggregate · Specs and context
- Prompts for work tasks · Prompts and review