Tutorials

Use Claude Code and Codex together

Tutorial: set up a Claude Code agent and an OpenAI Codex agent side by side, have one draft and the other critique, and understand how the two runtimes differ in sessions, approvals and cost.

#Goal

One Claude Code agent and one Codex agent in the same channel, each signed in with your own account, working on the same question.

#Prerequisites

LoCrew running.
Claude Code installed and signed in (claude).
Codex signed in: run npx codex from the LoCrew folder and sign in.

#1. Check both runtimes

Open Help → Restart setup if you want the runtime check screen, or simply start creating agents: the Create agent wizard checks each CLI when you choose it.

Expected: Claude Code found and Codex found. If Codex says found, not signed in, run npx codex again from the LoCrew folder.

Note

The wizard's check sends one tiny prompt to confirm the sign-in works, so it uses a very small amount of credit.

#2. Create the agents

	Agent 1	Agent 2
Name	Claude	Codex
Runtime	Claude Code	OpenAI Codex
Instructions	Propose an approach and explain trade-offs.	Critique proposals: find weaknesses and missing cases.
Working directory	Your test project	The same folder
Workspace access	Read only	Read only

Read only keeps this tutorial safe. For Codex in particular, LoCrew does not show its own approval dialog, so Read only is the safest choice when you are not ready for it to write.

#3. Ask both

Create a channel #design with both agents and send:

text

@all How should we cache user profiles in this project? Claude proposes, Codex critiques.

Expected: both agents are woken (@all addresses everyone) and reply in the same transcript. The status line shows Claude and Codex are working….

Behind the scenes: each agent's run starts in parallel, up to Max agents working at once (2 by default). The @all used two hops of the chain.

#4. Continue the conversation

Reply to just one: @Codex what would break first under load?. Only Codex is woken, and it resumes its own thread for #design.

#How the two runtimes differ

	Claude Code	Codex
Memory	A Claude Code session per conversation	A Codex thread per conversation
Instructions	Added to the system prompt every run	Sent at the start of a new thread only
Ask first	LoCrew's diff approval dialog	Codex's own approval policy; no LoCrew dialog
Cost	Real dollars reported by the CLI	Estimated from tokens and a price table
Turn limit	Enforced	No evidence of enforcement in the code
Verification	Tested against the real CLI	Install detection in the default suite; beta

See Runtimes for the full comparison.

#Common mistakes

Signing in to Codex in a different place than the CLI LoCrew uses. Run npx codex from the LoCrew folder.
Expecting Codex to show diffs for approval. Use Read only or a disposable copy of your project.

#Next steps

Sessions: why each agent remembers each channel separately.
Build a development crew