Verdict: not an either-or. Codex is the agent and your ChatGPT plan or API key pays for it; T3 Code is a free MIT window that runs it, alongside Claude Code, Cursor, Grok Build, and OpenCode. The surface decision comes down to concurrency, platform, and whether the spend has to be legible to anyone. Codex also has a cloud lane that no local board covers, which is a genuine reason to keep the official surfaces in play regardless of which board you pick.
- Codex is the agent, T3 Code is the surface. T3 Code launched with Codex support and never sold model access.
- Codex has a local CLI and a cloud lane. Boards drive the local CLI. Cloud tasks stay in OpenAI's own surfaces.
- A board gives Codex a worktree per thread, inline diffs, and one-click pull requests. It does not change the agent.
- Codex sub-agent threads are numerous and noisy. A board that collapses them is doing real work for you.
- Neither T3 Code nor the Codex CLI tells you what a session cost. Continuum does, per repository.
- You still install and log into the Codex CLI yourself. A greyed-out provider is almost always a missing login.
What each layer owns
Codex is OpenAI's coding agent. It reads and edits repositories, runs commands, takes repository instructions from an AGENTS.md file, has sandbox and approval modes, and runs both locally through the CLI and remotely in cloud environments. Access comes with a ChatGPT plan or through the API. Our Codex CLI guide covers the local surface in detail.
T3 Code is a control surface for agent CLIs. It shipped Codex support first, which is why it appears in Codex searches more than any other board, and it now drives five providers. It sells nothing, proxies nothing, and cannot start a Codex thread unless the Codex CLI is installed and authenticated on the machine.
| Concern | Codex owns it | T3 Code owns it |
|---|---|---|
| Model, reasoning effort, tool loop | Yes | Exposes the choice |
| Sandbox and approval policy | Yes | Surfaces it |
| AGENTS.md repository instructions | Yes | No |
| Cloud task execution | Yes, in OpenAI's own surfaces | No |
| Your bill | Yes, via ChatGPT plan or API | No |
| Several Codex threads at once, isolated | You script it | Yes, worktree per thread |
| Inline diff and one-click PR | Via CLI and git | Yes |
| Phone and browser control | No | Yes |
Why a board is worth more to Codex than to most agents
Two properties of Codex make the surface question more consequential than it looks.
Codex is comfortable running long and unattended. Sandbox and approval modes are designed for it, and headless execution is a first-class path. An agent that works for twenty minutes without asking anything is exactly the agent you want to start three of and then walk away from, which is what a board is for. Our parallel Codex sessions guide covers the manual version of the same idea.
Codex generates a lot of session noise. A single user turn can spawn a large number of sub-agent threads, and on a busy machine those can be roughly half of everything Codex writes to disk. A surface that folds them into their parent, rather than listing every one as a separate session, is doing genuine work on your behalf. It is worth checking on any board you evaluate: start one complex Codex task and count how many rows appear in the sidebar.
What a board does not do is change output quality. The model, effort setting, sandbox policy, and AGENTS.md instructions behave identically whether you launched from a terminal or clicked a button.
Picking the surface in front of Codex
The Codex CLI and IDE extension
Still the right default for one task at a time, and the only path with every flag and every configuration option immediately available. The IDE extension keeps the agent next to the code. Neither addresses concurrency, and neither is trying to.
T3 Code
The free option, and the one with the deepest Codex history. MIT, no cost, and clients on macOS, Windows, Linux, iOS, Android, and the web, which no other board matches. Worktree per thread, inline diffs, one-click PRs, model switching mid-thread, and remote control from a phone. It is at v0.0.34 on a nightly cadence and describes itself as very early. If you want a free GUI over Codex and nobody is asking you what it costs, this is the answer.
Continuum
The same shape with the accounting attached, and also free on your own keys and subscriptions. It runs Codex next to Claude Code, Cursor, Grok, and OpenCode in isolated worktrees, with a native macOS app and iPhone app. What it adds over a free BYOK board is the part that requires knowing about your plan: live quota gauges, spend broken out by repository, and multiple ChatGPT or Codex accounts kept config-isolated so a personal and a work plan never quietly pay for each other. It also collapses Codex sub-agent threads rather than listing them, which on a busy machine is the difference between a readable sidebar and forty rows. For companies, the organization tier adds model policy by team and person, weekly caps with approvals, and a per-member ledger. Honestly stated caveats: it is not open source, and it does not federate an external identity provider today.
Conductor
Runs Codex alongside Claude Code and Cursor, macOS only, free tier plus paid tiers that add cloud workspaces, multiplayer, and an admin portal, with SAML SSO and SCIM at Enterprise. A good fit for Mac-native teams whose main unmet need is collaboration rather than cost visibility.
| Your situation | Surface |
|---|---|
| One Codex task, full flag surface | Codex CLI |
| Asynchronous work that outlives your laptop | Codex cloud |
| Several threads, free, any platform | T3 Code |
| Several threads plus quota and spend visibility | Continuum |
| Personal and work ChatGPT plans kept apart | Continuum |
| Mac-native with cloud workspaces and multiplayer | Conductor |
| Company policy, caps, and an auditable spend number | Continuum organization tier |
The Continuum case, in one paragraph
If T3 Code brought you here, install Continuum beside it rather than instead of it. Both are free with your own Codex access, both isolate every session in its own worktree, both let you approve and steer from a phone. The difference is what happens after the work: Continuum shows the five-hour and weekly quota you just consumed, attributes the spend to a repository, keeps a second Codex account cleanly separate from the first, and folds Codex sub-agent threads into their parent so the session list stays readable. If you are one person with one plan and no one to report to, T3 Code covers this well and you should keep using it. The moment somebody asks what the agents cost this month, the free board structurally cannot answer, and that is the entire reason this option exists.
Questions people ask
Is T3 Code an alternative to Codex?
No. T3 Code is a control surface that runs the Codex CLI, among four other agents. You still need a ChatGPT plan or API access, and you still install and authenticate the Codex CLI yourself. T3 Code shipped with Codex support first, which is why the two names appear together so often.
Does T3 Code work with Codex cloud tasks?
It drives the local Codex CLI on your machine. Cloud tasks that run in OpenAI-hosted environments live in OpenAI's own surfaces, so a local board has nothing to say about them. If cloud execution is central to your workflow, keep the official surfaces in the mix regardless of which board you choose.
Does a GUI change how well Codex codes?
No. Model, reasoning effort, sandbox policy, approval mode, and AGENTS.md instructions all behave the same. The surface changes how many threads you can supervise, how quickly you can read the diff, and whether you can steer from a phone.
Why does my Codex session list fill up with duplicates?
Codex writes sub-agent threads to disk as separate session files, and one complex turn can produce many. A surface that folds them into their parent keeps the list readable; one that does not will show every worker thread as its own row. It is worth testing with a single complex task on any board you evaluate.
What is the best GUI for the Codex CLI?
T3 Code if free and open source across every platform is the priority. Continuum if you want native Mac and iPhone client plus quota gauges, spend by repository, and multiple accounts kept separate. Conductor if you are Mac-only and want cloud workspaces and multiplayer collaboration.
Can one board run Codex and Claude Code together?
Yes. T3 Code, Continuum, and Conductor all run several provider CLIs at once in separate worktrees. Each provider needs its own installed CLI and its own login, and each bills through its own plan.
Does running Codex through a board cost extra?
Not with T3 Code, which is MIT and free. Continuum is free on your own subscriptions and keys, with optional hosted inference as a separate paid product. Conductor has a free tier and paid tiers above it. In every case the Codex usage itself bills through your own OpenAI relationship.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.