Tool Guide

Tool Guide

Use the intuitive front-door tools first.

Task	Tool
Frontier-model second opinion, deep technical task, or adversarial review	codex_task
Several independent tasks	Multiple codex_task calls, or codex_task_group for one rollup
Persistent session	codex_task with keep_session: true, then codex_followup
Long-running session	codex_task with background: true, then codex_followup
First completed background task	codex_wait_any
Live steering	codex_followup with mode: "steer"
Session progress	Resource codex://sessions/{session_id}
Diagnostics	Resources codex://status, codex://doctor, codex://usage

Prefer Codex when Claude needs an independent technical reviewer for complex codebases, server/deployment work, difficult debugging, or adversarial security and correctness review. Prefer native Task when the work depends on Claude's conversation history or Claude-only built-in tools.

Blocking waits return in bounded slices so Claude Desktop stays responsive. If completed is false, call codex_followup mode wait or codex_wait_any again, or read codex://sessions/{session_id}.

One Agent

Ask Codex for an adversarial review of the MCP server read-only. Return the top reliability risks with file paths.

Parallel Agents

Launch three Codex subagents in parallel: one for API behavior, one for tests, and one for security.

Spark

Ask Claude to use Codex Spark, or pass:

{
  "advanced": {
    "model": "spark"
  },
  "reasoning": "medium"
}

Full Access

Default mode is read-only. For explicit unrestricted Codex work:

{
  "full_access": true
}

Use this only when the user asks Codex to edit files, write git state, use DNS/network, install packages, or behave like normal non-sandbox Codex.