This page covers the AI tools used in daily development and when to reach for each one.

​
Tool summary

ToolRoleBest for
Claude (web)Research and reasoningExploration, drafting, design decisions
Claude Code (CLI)In-terminal agentCode editing, file operations, git workflow
OpenClawOrchestratorPlanning, routing, task coordination
CodexImplementerFocused code execution, repo inspection

​
Claude

Claude at claude.ai is the starting point for tasks that require reasoning, research, or generating options before committing to a path. Use Claude for:
  • Architecture decisions and trade-off analysis
  • Drafting documentation and content
  • Code review when you want a second read without touching files
  • Evaluating AI tool updates before adoption
Avoid using Claude for:
  • Tasks that need direct file access — use Claude Code instead
  • Anything requiring a multi-turn execution loop — use OpenClaw

​
Claude Code

Claude Code is the CLI version of Claude with full file system access, shell execution, and git integration. It runs directly in the terminal against your codebase. 
# Start a Claude Code session
claude

# Or use it inline
claude "fix the type error in tools/dev-environment.mdx"
Use Claude Code for:
  • File editing tasks with a clear scope
  • Running commands and reading their output
  • Making targeted changes across multiple files
  • Reviewing diffs and git state
Key settings: Claude Code respects .claude/settings.json and CLAUDE.md files in the project root. Set permissions and tool allowlists there to reduce prompts.

​
OpenClaw

OpenClaw is an orchestration layer that sits above execution agents. It handles planning, context management, routing, and quality control without touching files directly. Use OpenClaw for:
  • Scoping and breaking down complex tasks
  • Routing subtasks to the right execution agent
  • Defining done criteria and quality gates
  • Reviewing execution output and deciding next steps
  • Managing multi-step workflows where context needs to persist
OpenClaw does the thinking. Codex does the doing.

​
Codex

OpenAI Codex (via the Codex CLI) is the focused implementation agent. It inspects repos, makes targeted code edits, runs tests, and reports results. 
codex "implement the delivery checklist component in the UI"
Use Codex for:
  • Implementation tasks with a locked, narrow scope
  • Running and interpreting test output
  • File-level edits where the outcome is clearly specified
Avoid using Codex for:
  • Exploratory or research tasks
  • Tasks without clear done criteria — it will fill ambiguity with plausible-looking but wrong code

​
Choosing between Claude Code and Codex

Both run against your codebase. The difference is context and oversight:
DimensionClaude CodeCodex
OrchestrationSelf-directedDirected by brief
Context windowLong, persistent per sessionNarrow, scoped to task
AutonomyHigher — works through ambiguityLower — needs explicit scope
Best whenExploratory or multi-stepExecution of clear, bounded tasks
In the OpenClaw + Codex workflow, OpenClaw uses Claude Code’s reasoning for planning and routes implementation to Codex. See OpenClaw + Codex Workflow for the full pattern.

MCP Servers

Extend Claude Code with MCP servers for filesystem, search, and external integrations.