Codex CLI in NeuroSquadCodex CLI,on a canvas of its own
Run as many Codex sessions as you need, each in its own card with a real terminal. NeuroSquad knows when each one is working, waiting for your approval or finished, hands it browsers, terminals and notes by arrow, and picks up every conversation where it left off.
Official sitedevelopers.openai.com/codex/cli- The real
codexCLI, your own login - Nothing written to
~/.codex - Free, for Windows 10 and 11
It knows when Codex is done — and when it’s asking
A quiet terminal can mean “finished” or “waiting for your approval”. NeuroSquad doesn’t guess: every Codex card reports its own turns through Codex’s lifecycle hooks.
UserPromptSubmithookYou press Enter. The card’s icon spins and the Squad status card counts it as working. Subagents Codex starts don’t move the card’s status.
PermissionRequesthookCodex stops at its approval prompt. The card pulses, a sound plays once, and the toast leads with what it wants to run — Bash — rm -rf dist. A question from request_user_input counts too.
StophookThe turn is over — not just quiet. The next queued prompt goes out, the journal gets the answer, and a toast says it’s done. Esc mid-turn ends it through the Interrupt hook.
A failed turn still ends
Codex runs no hook after an error. While a turn is open, its own “■ …” error line closes it, so a card never hangs on “working”.
Approvals by a reviewer aren’t pings
If you route approvals to Codex’s automatic reviewer, nobody is asked — so the card doesn’t call you either.
Trusted, not bypassed
The hooks are passed with their trust hashes, so there is no review screen at startup — and your repo’s own untrusted hooks stay untrusted.
Three things you’ll do every day
Replicas of the app, played step by step. When Codex asks for approval, the answer is yours.
Codex fixes a failing test, stops at its approval prompt before deleting a folder, and finishes — the journal writes itself.
Everything Codex CLI can do in NeuroSquad
Codex is integrated as fully as Claude Code — through its own hooks and session flags. Here is all of it, with a docs page for each, and the few things Codex itself doesn’t allow.
- 01
Knows its state
Codex’s own hooks report every turn, so nothing is guessed.
Finished and waiting are different
Each card gets Codex’s lifecycle hooks:
Docs: Finished and waiting are differentUserPromptSubmitmarks it working,PermissionRequest— waiting on you,StopandInterrupt— finished.A notification that says what it needs
The toast leads with what Codex wants:
Docs: A notification that says what it needsBash — rm -rf dist, the files anapply_patchtouches, or the MCP tool. One per agent; a waiting one stays until you answer; a click flies to the card.Picks up where it left off
Every hook reports the card’s session, so the next launch is
Docs: Picks up where it left offcodex resume <id>in the card’s own folder. A session whose file is gone starts fresh by itself; “Restart session” gives a clean one.The Squad status card
Every AI agent of the workspace on one card — waiting on you, working, finished, idle — with the model from Codex’s own session log (or your
Docs: The Squad status cardconfig.toml) and the live time of its turn.Custom cards can follow it
Cards built with the Card SDK see the same statuses and turn events, and read Codex’s last answer from its session file.
Docs: Custom cards can follow it
- 02
Works through the canvas
An arrow from the card is a set of tools. Draw it, and Codex can use what is at the other end.
Arrows are tools
An arrow to a browser, terminal, note or another agent gives Codex
Docs: Arrows are toolsbrowser_*,terminal_*,note_*oragent_*over NeuroSquad’s MCP server, approved without a prompt — the arrow is the consent. Each call lights up its arrow.It can add cards itself
Docs: It can add cards itselfcanvas_spawn_cardlets it open a terminal, a browser, a note or a helper agent — Codex or any other CLI — next to itself, already joined by an arrow.Canvas mode
Codex starts with
Docs: Canvas modefeatures.shell_tool=falseandweb_search=disabled: its own shell and web search aren’t offered at all, so even dangerous mode can’t bring them back. Commands run in terminal cards you can watch.MCP servers from the registry
Install a server from the official MCP Registry once, then give it to an agent with an arrow. Codex asks before every call to it, and its secrets never leave NeuroSquad.
Docs: MCP servers from the registrySkills from skills.sh
An arrow to a skill card gives Codex
Docs: Skills from skills.shskill_loadwith each skill and when to use it. A new skill arrow reaches a running Codex after a card restart.
- 03
Built for long runs
Long sessions run into full contexts and usage limits. Both are handled on the card.
Context meter and Compact
A small ring in the header shows how full the context is — the last request’s tokens against the model’s window, both from Codex’s own log. Compact sends
Docs: Context meter and Compact/compact.Continues after a usage limit
When Codex says “You’ve hit your usage limit … try again at 3:00 PM”, the card counts down to that time and sends “continue” then. Off per card if you prefer.
Docs: Continues after a usage limitPrompt queue
Line up the next messages: each goes out when the turn is finished — with the hooks, “finished” is exact. Daily prompts too.
Docs: Prompt queueHand off to a fresh agent
When a session is full, start a new agent beside it — Codex or any other CLI — with a summary of the work typed in, ready for you to send.
Docs: Hand off to a fresh agentJournal
Each finished answer is added to a connected note, read from Codex’s session file: a running log of what the agent did.
Docs: Journal
- 04
Works in a team
Several Codex agents — or a mix of CLIs — on one project without stepping on each other.
Isolated worktrees
Give an agent its own git worktree, so parallel agents never edit the same files. The folder-trust setting follows the worktree, so Codex opens straight to the prompt.
Docs: Isolated worktreesAdd a reviewer
One click starts an agent on a different CLI in the same worktree, with an arrow from the author and a prompt to review without editing anything. Codex is also the first choice when another CLI needs a reviewer.
Docs: Add a reviewerDangerous mode, per card
For a run you trust, a card can start Codex with
Docs: Dangerous mode, per card--dangerously-bypass-approvals-and-sandbox. It’s a per-card switch, off by default.Roles
Reviewer, tester, researcher, architect, writer or your own: the role is typed in as a first prompt you can see and edit, and never re-sent behind your back.
Docs: RolesAGENTS.md, natively
Codex reads AGENTS.md itself. Edit it once, and NeuroSquad mirrors it for the other CLIs on the project.
Docs: AGENTS.md, natively
- 05
Cost, control and access
What it spends, what it may spend, which model it runs on, and how you reach it.
Exact usage and cost
Tokens and cost by card, workspace and model, read from Codex’s own session logs — a subagent’s requests fold into the session that started it. Whole tokens, exact costs.
Docs: Exact usage and costBudget brake
A limit per workspace. Past it, each Codex card gets Escape — the turn stops, the session is kept — and automatic prompts stop. Raise the limit and it’s lifted.
Docs: Budget brakeOpenRouter models
Run a card on an OpenRouter model instead of your OpenAI login. The key goes only into that card’s environment, and the session stays resumable even when the key is gone.
Docs: OpenRouter modelsDictation into the prompt
Hold a hotkey and speak. Speech is recognised on your machine and typed into the active agent’s prompt.
Docs: Dictation into the promptFrom your phone
Scan a QR code and the whole canvas opens in the phone’s browser: read Codex’s terminal, see what it asks, send the next prompt.
Docs: From your phone
- 06
What Codex doesn’t allow
Where Codex itself has no way to do what Claude Code does, and what NeuroSquad does instead. Checked in Codex’s source.
No session id chosen up front
Codex generates its own thread ids and has no flag to start with one you pick. So the card learns its id from the first hook and resumes by it — a card that never ran a turn simply starts fresh.
Docs: No session id chosen up frontNew skill and MCP arrows need a restart
Codex lists a server’s tools once and ignores later changes. NeuroSquad offers every built-in tool from the start and lets the arrows decide at call time — but an arrow to a skill or MCP-server card reaches Codex on its next start.
Docs: New skill and MCP arrows need a restartCtrl+C would quit it
On an idle Codex, Ctrl+C exits the program. So the budget brake presses Escape instead, which interrupts the turn and leaves Codex running.
Docs: Ctrl+C would quit itNo attribution via the built-in provider
If you point Codex’s built-in
Docs: No attribution via the built-in provideropenaiprovider at OpenRouter yourself, Codex takes no extra headers for it. Your own custom providers for OpenRouter do get them.“Answered” is read from Enter
No Codex hook reports that you answered an approval, so the card goes back to working when you press Enter. Answer with the
Docs: “Answered” is read from Enteryshortcut and it stays “needs you” until the turn ends.
Who’s waiting, and what it cost
Two cards that answer the questions you ask most when several Codex sessions run at once.
The Squad status card
Every agent of the workspace with its status, model and turn time. The one waiting on you comes first, with what it wants to run. Tap a status to filter.
- coupon-fixCodex CLIgpt-5.5Bash — rm -rf distNeeds you2:34
- api-migrationCodex CLIgpt-5.5Move routes to the new clientWorking14:58
- db-indexesCodex CLIgpt-5.4-miniAdd the missing indexesWorking6:42
- docs-passCodex CLIgpt-5.4-miniUpdate the READMEFinished4 min. ago
Usage & cost
Read from Codex’s own session logs on your disk and matched to the card by the sessions its hooks reported. The rows always add up to the total.
| Agent | Tokens | Cost | ||
|---|---|---|---|---|
| coupon-fix | 1,196,402 | $2.74 | ||
| api-migration | 781,355 | $1.63 | ||
| docs-pass | 238,610 | $0.09 | ||
| Total | 109 | 2,216,367 | $4.46 |
The real Codex, plus a few session flags
Codex has no per-run settings file, but any setting can be given for one run with `-c`, over your `config.toml` and never written into it. This is everything NeuroSquad adds:
Nothing written to ~/.codex
Everything is a -c flag for this run. The MCP token is passed in the environment, never in the arguments.
Your login, your plan
Codex runs with the ChatGPT sign-in or API key it already uses. Your NeuroSquad account is separate and never sees it.
The real codex.exe, directly
npm’s codex.cmd only starts a script that starts codex.exe — the card starts the exe itself, so stopping the card really stops Codex.
Codex CLI in NeuroSquad, answered
Give Codex a canvas
Free for Windows 10 and 11. Your code, your login and your costs stay on your machine.