Devin vs Claude Code Comparison

# Devin CLI — install, choosing a permission mode, and session management
# Source: https://docs.devin.ai/cli and https://docs.devin.ai/cli/reference/commands

# 1) Install Devin CLI (macOS/Linux)
curl -fsSL https://cli.devin.ai/install.sh | bash

# 2) Start a task in sandboxed Autonomous mode inside a project directory
cd ~/projects/backend
devin --permission-mode autonomous --sandbox \
  -- "Upgrade Express 4 routes to Express 5, keep the test suite green"

# 3) Resume the latest session in the same directory, or return to a specific session
devin -c
devin -r abc12345

# 4) A one-off query in non-interactive (print) mode
devin -p "list all TODO comments"

# 5) Start a cloud session instead of the local agent
devin --cloud

# Switch modes mid-session: /mode, /normal, /accept-edits, /smart, /bypass
# Note: Autonomous only runs with --sandbox, and since edit/write calls fall
# outside the sandbox, it still asks for approval in this mode too;
# the mode that runs every call without approval is Bypass.

# Claude Code — starting a task in sandboxed Auto mode and tracking cost
# Source: docs.claude.com/en/docs/claude-code/setup · /security · /costs · /github-actions

# 1) Install (native installer — macOS/Linux/WSL)
curl -fsSL https://claude.ai/install.sh | bash

# 2) Start interactively in Auto mode inside the repo
cd ~/projects/backend
claude --permission-mode auto \
  "Upgrade Express 4 routes to Express 5, keep the test suite green"

# 3) In-session token/cost summary
/usage

# 3b) The budget cap only applies in print (headless) mode — -p is required
claude -p --max-budget-usd 5.00 \
  "Upgrade Express 4 routes to Express 5, keep the test suite green"

# 4) Trigger from a GitHub Actions PR/issue comment (install first)
/install-github-app
# Then, in a PR/issue comment: @claude review this PR and list edge cases

# Note: in Manual mode every file write/command run requires approval;
# in Auto mode a separate classifier model reviews risky actions.