Skip to main content
Devin Local sessions offer five modes. Each mode sets what the agent may do and how much it asks before acting. Switch modes with the mode selector below the input box. Because Devin Local shares its agent harness with Devin CLI, the modes map onto the Devin CLI agent modes and permission modes; the slash commands listed in essential commands also work in the Devin Local input. Whatever the mode, your permission rules and any organization-enforced rules still apply. Deny and ask rules your admin enforces take priority even in Bypass Permissions.

Code mode

Code mode is the default mode for making changes to your codebase. In Code mode, the agent can:
  • Create, edit, and delete files
  • Run terminal commands
  • Search and analyze your codebase
  • Install dependencies
  • Execute multi-step tasks autonomously
Edits inside the workspace are auto-approved. Shell commands, web fetches, and MCP tools prompt for approval unless an allow rule covers them. If your organization does not allow auto-approved edits, Code mode prompts for edits as well.
We recommend Code mode as your default for most tasks.

Smart mode

Smart mode keeps Code mode’s auto-approved edits and adds a fast model that decides, for every other action, whether it is safe to run without asking. Anything it does not judge clearly safe still prompts, and high-risk categories (package installs, mutating git, rm, sudo, destructive cloud CLI operations, sensitive files) always prompt.
Smart mode is rolling out gradually, so it may not appear in your mode selector yet. See Smart Mode for the full behavior.

Ask mode

Ask mode is a read-only mode optimized for questions and exploration. In Ask mode, the agent can search and analyze your codebase, but cannot make any changes.

Plan mode

Plan mode helps you think through complex tasks by developing a detailed implementation plan before writing any code. In Plan mode, the agent will:
  • Explore your codebase to understand the current state
  • Ask clarifying questions to ensure the plan aligns with your goals
  • Provide multiple options for you to choose from with an interactive interface
  • Present a detailed plan, written in an external Markdown file, with implementation steps
The plan file lives outside your repository and persists for the session, so planning can span several messages. When the agent is finished, it asks to start implementing — or you can click Implement on the plan file yourself to switch to Code mode and begin.

Planning more deeply

Including the keyword megaplan, ultraplan, or masterplan anywhere in your message switches the conversation into Plan mode and makes the agent plan more extensively, asking at least one clarifying question before it writes the plan. The previous mode is restored once you remove the keyword.

Continuing from a plan

The Markdown file created in Plan mode can be particularly useful for continuing work across multiple sessions. Plans are stored in ~/.devin/plans/ and are available in the @mentions menu. By mentioning a plan file, you can continue implementation with a fresh context. This can be particularly useful when an initial implementation went awry: discard the original changes, tweak the plan file, and click Implement to attempt implementation again in a new conversation.

Exiting Plan mode

There are multiple ways to move from planning to implementation:
  • Click the Implement button on the plan file
  • Approve the agent’s request to leave Plan mode once the plan is ready
  • Change your mode to Code mode in the input box
  • Let the agent automatically switch to Code mode when it detects that you’re ready to implement

Bypass Permissions

Bypass Permissions auto-approves all tool calls, including file writes and shell commands.
Use Bypass Permissions only for work you fully trust. It never overrides organization-level permissions configured by your admin: admin-enforced deny and ask rules always take priority.

Sandboxed sessions

When sandboxing is enabled, the session runs in Autonomous mode instead of the modes above: the OS-level sandbox contains what commands can touch, and you are prompted for capabilities (such as network access) rather than individual commands. See Autonomous Mode.