Compare·Provider
Continuum
Multi-agent workbench
VS
Codex
OpenAI coding agent

Codex in Continuum

Same Codex runtime, same OpenAI bill. Continuum adds the peers, the worktrees and repo-level cost truth.

Codex is OpenAI's coding agent, reachable as of August 2026 from the CLI, the IDE extension, the web, the ChatGPT desktop app and iOS, running either locally in a sandbox or in a cloud environment you delegate to. Continuum is the multi-provider control plane that hosts Codex as a first-class agent beside Claude Code: a git worktree per session, six review panes on one thread, isolated CODEX_HOME roots for work and personal auth, live quota gauges, and a local dollar ledger. Continuum does not replace Codex; it drives it. Live Codex chat runs through Continuum's ACP harness on codex app-server, which is what makes tools, turn state and completion signals render honestly instead of being guessed from a log file.

Download for Mac All comparisons
Workflow Monday path Matrix Decide Money FAQ

Updated 2026-08-03 · Mac stable · Win/Linux desktop beta

Choose Continuum when

You want Codex and Claude Code in one sidebar, isolated work and personal Codex logins, worktree sessions with diff and PR panes, quota gauges across accounts, and approve or interrupt from the phone.

Choose Codex when

You want OpenAI's native Codex experience and nothing around it: the CLI, the IDE extension, the web and cloud tasks, all inside one vendor's product surface.

Snapshot provider · Continuum hosts it
Dimension Continuum Codex
Engine identity Hosts official Codex via ACP harness Is OpenAI's coding agent
Auth path Your ChatGPT or Codex auth · no markup codex login on your OpenAI plan
Accounts per machine Several · isolated CODEX_HOME roots One default auth home
Other providers Claude Code · Cursor · Grok · Gemini · OpenCode OpenAI stack only
Surfaces Mac · web · Win/Linux · iPhone · Watch CLI · IDE extension · web · desktop · iOS
Quota gauges Menu-bar 5h + weekly, per account Limits per 5h window, no fleet gauge
Analytics Local JSONL → $ by repo/provider/day Account usage, not a repo ledger
Mac stable · web · iPhone · Watch · Win/Linux desktop beta · free app
01

OpenAI silo vs multi-provider switchboard

In Continuum

Codex in a multi-provider day

09:30

Spawn a Codex session in the Code tab. The harness streams turns, tool calls and edited-file counts into the same thread as the plan and the diff, so nothing has to be reconstructed from a log.

11:00

Open a Claude Code session on the same prompt in a second worktree. Compare the diffs side by side at review and keep the stronger branch; the other one is archived, not merged out of politeness.

14:20

A plan needs a decision while you are in a meeting. Approve it from the iPhone; the send goes through an idempotent outbox, so tapping twice on bad wifi cannot double-post.

18:45

The work CODEX_HOME stays isolated from personal auth, and the night batch runs pinned to it. Analytics still attribute both to the right provider column and the right repo.

In Codex

Native Codex only

09:30

Use the Codex CLI, the IDE extension or the web. Full OpenAI agent quality, native sandboxing, and no third-party layer between you and the runtime.

11:00

Push a longer job to a cloud environment and let it run without pinning your laptop. That is a genuinely good answer when the work does not need your local toolchain.

14:20

Comparing the answer to Claude means opening another product. There is no second provider in this pane, and OpenAI has no reason to add one.

18:45

One default auth home. Work and personal on the same machine means logging out and back in, and the usage view is per account, not per repo.

02

Monday path

Monday with Codex in Continuum's Code tab

Install, connect, first session - steps you can run the same day.

01

Install Continuum and confirm codex logs in normally first. A broken CLI login is still a broken login inside a workbench.

02

Code tab, new session, Codex agent, worktree mode for the ticket. Chat, plan, diff, PR, terminal and artifacts panes attach to the one thread.

03

Open a Claude Code peer session on the same repo when you want a second line of attack, then compare both diffs at review instead of trusting one.

04

Pair the iPhone so you can approve or interrupt while Codex keeps running on the Mac or on an enrolled Linux host.

05

Add the work account as a second instance. Continuum gives it its own CODEX_HOME so the personal auth.json is never stomped.

06

Check Usage for Codex history by repo and provider, and the gauges for the windows Continuum can poll. Background: Codex pricing.

07

Before a long ACP run, confirm which account the session is pinned to. Continuum refuses to spawn on an unresolvable pin rather than quietly billing the wrong plan.

08

For a spike you do not want on your laptop, delegate it to a Codex cloud task from the session terminal and keep the local worktree for the work you are reviewing.

How Continuum actually runs Codex

How Continuum actually runs Codex

  • Live Codex chat drives codex app-server through Continuum's ACP harness. Nothing about the model is reimplemented, and the transcript is not scraped from a rollout file.
  • Turn state, tool calls, edited-file counts and completion notifications come from harness events, so a long turn reads as running instead of silently looking finished.
  • Each managed session can own a git worktree, so a Codex run and a Claude run on the same repo never contend for one working tree.
  • Secondary accounts get their own CODEX_HOME under a Continuum instance root, and the primary account's spawn environment stays byte-identical to the bare CLI.
  • The session file resolver matches a rollout's recorded working directory before adopting it, so a stray codex in another terminal cannot hijack a live transcript.
  • Analytics parse Codex session JSONL into the same local ledger as Claude, priced with a LiteLLM snapshot, so one number covers both providers per repo.
  • The spawn grid can open plain interactive codex tiles, four, six or eight at a time, when you want raw TUI beside the managed worktree sessions.
03

Codex host matrix

Cells use product nouns on both sides, checked August 2026. Same runtime under your own OpenAI auth, so nothing here re-scores model quality. Continuum cells name the cockpit: worktrees, panes, account isolation, gauges and the ledger. Codex cells name what OpenAI already ships, including cloud delegation and local sandboxing.

Capability Continuum Codex
Official Codex agent ACP harness on codex app-server Is Codex
Local sandboxed execution Codex sandbox + worktree isolation Native sandbox, incl. Windows
Delegate a task to the cloud Via Codex cloud from the session Native cloud environments
Claude Code in the same app First-class peer, same sidebar OpenAI stack
Parallel multi-provider sessionsOne ticket, two engines, two worktrees Fleet sidebar + worktrees Multiple Codex tasks
Two Codex logins on one machine Isolated CODEX_HOME per account One default auth home
Worktree Code panes Chat · plan · diff · PR · terminal · files Agent-native surfaces
Hunk-level diff review before PR Stage or revert per hunk, then gh Your own git tooling
Local $ and token ledger JSONL priced by repo and day Account usage in ChatGPT
iPhone and Watch control Approve, interrupt, diff, PR Codex on iOS for its own tasks
Live quota gauges across accounts Menu bar + Usage tab Limits shown in the product
Spawn grid of interactive tiles 4, 6 or 8 codex TUI tiles Your own terminal windows
Run on Linux or Windows hosts you enrol Go agent host + same panes CLI runs where you install it
Cost of the workbench itself Free · optional hosted inference Included with your ChatGPT plan
04

What Continuum adds around Codex

01 · Single vs multi

Codex alone is a silo. Continuum is a switchboard.

If every session is Codex, OpenAI's own surfaces may be all you need. The moment you also pay for Claude Max, or want two engines to attack the same ticket, Continuum becomes the place the sessions live. It also hardens what a session is allowed to adopt: a rollout must match the session's own working directory before it counts, so a stray terminal cannot hijack a live transcript.

02 · The workbench around the runtime

Transcript, plan, diff and PR on one thread.

Continuum's Code tab is built for review, not just for watching an agent type: a plan tracker, a git diff with per-hunk staging, a PR pane wired to gh, artifacts with QuickLook, and multiple terminals. Codex is the runtime; this is how you operate it for eight hours without losing the thread. Related: reviewing AI-generated code.

03 · Accounts

Work Codex and personal Codex without clobbering auth.

Continuum isolates CODEX_HOME per account under its own instance root, so a second login never stomps the first, and the primary's environment stays identical to the bare CLI. Surgical isolation, not a HOME swap, because overriding HOME breaks git, ssh, gh and node inside a worktree.

04 · Cost truth

Repo-level dollars, not vibes.

Continuum aggregates Codex and Claude history into one local ledger and prices it with a LiteLLM snapshot, so you can see which repo and which provider ate the week beside the live windows. OpenAI shows you the account. Only you can see the repo. See cost tracking.

05 · Which path is live

Harness, not log scraping.

Live Codex chat in Continuum is the ACP harness path. The historical JSONL decoders are analytics and history only, which is why a fix aimed at Codex chat behaviour belongs in the harness. It matters to users because turn state, tool output and completion notifications are only honest when they come from the runtime itself.

05

Coexistence

Stack recipe

Same engine. Bigger cockpit.

Stack recipe: Continuum drives the real Codex through the ACP harness on codex app-server, under your own ChatGPT or Codex auth, so usage stays on your OpenAI account and Continuum takes no cut. Keep OpenAI's native surfaces for OpenAI-only days and for delegating a task to a cloud environment. Open Continuum when Claude peers, two Codex logins on one machine, worktree isolation or phone approve matter. Compare the engines themselves in Codex vs Claude Code.

Download Continuum Providers
06

Host Codex or stay native-only

Continuum

Choose Continuum when Codex is one peer among many

  • Run real Codex beside Claude Code and other agents in one sidebar, on one thread each.
  • Isolate work and personal Codex logins with separate CODEX_HOME roots on the same machine.
  • Give every session its own git worktree so parallel runs never fight over the working tree.
  • Approve, interrupt and review diffs from iPhone or Watch while the host keeps running.
  • See local dollars and tokens by repo for Codex and Claude in one ledger. Try parallel Codex sessions.
  • Keep the workbench free on top of the ChatGPT plan you already pay for.
Codex

Stay native when OpenAI alone is the whole stack

  • Only want OpenAI's native Codex product surface, end to end.
  • Do not need Claude or other labs in the same workbench.
  • Prefer delegating long jobs to Codex cloud environments over hosting them yourself.
  • Are happy with one auth home and OpenAI's own usage view.
  • Do not need a phone client that steers sessions on your own hardware.
  • Want the smallest possible stack: Codex, alone.
07

What you pay

Continuum

App + your labs

$0 for the app on Mac, web, Windows, Linux, iPhone and Watch. Bring your own keys, or the Claude and ChatGPT plans you already pay for, and Continuum takes no cut of those sessions. As of August 2026, optional Continuum-hosted inference is Plus $25/mo, Max 100 $100/mo, Max 200 $200/mo, or Ultra $500/mo. Codex sessions always bill your OpenAI account.

Codex

Their bill

Included with your ChatGPT plan as of August 2026: Free for quick tasks, Go $8/mo, Plus $20/mo, Pro $200/mo with maximum Codex tasks, Business $20/user/mo on annual billing or $25 monthly, and Enterprise or Edu by sales. Limits are per 5-hour window and vary by model; ChatGPT credits extend them. An API key is the pay-per-token alternative. See Codex pricing.

How to compare

Total cost of work

Continuum's harness runs Codex on your own auth, so there is no Continuum markup on those tokens. The only Continuum line is optional hosted inference. Compare the whole stack in AI coding pricing compared.

Codex plans checked 3 August 2026 against their pricing page. Vendors reprice often - if you spot a stale figure, tell us and we will fix it. Continuum never marks up a session it runs under your own login.

Continuum ladder: Free app · Plus $25/mo · Max 100 · Max 200 · Ultra - full pricing. Also analytics, multi-account, devices.

08

Questions

Codex + Continuum.

Deep dives: docs, providers, sessions.

Wrong comparison: Continuum runs Codex. It is better when you need multi-provider sessions, isolated accounts, worktrees, mobile control or a local ledger. Codex alone is better when you want OpenAI's native product with nothing else in the way.

Free covers quick tasks. As of August 2026 the paid rails are Go at $8/mo, Plus at $20/mo, and Pro at $200/mo with maximum Codex tasks, with Business at $20/user/mo annual. An API key is the pay-per-token route. Continuum does not replace that bill. See Codex pricing.

No. Live Codex chat drives codex app-server through Continuum's ACP harness, so the agent, the sandbox and the model are OpenAI's. Continuum owns the projection into panes, the account isolation and the ledger.

Yes, and it is the main reason people install Continuum. Open parallel sessions in separate worktrees, let each finish, compare the diffs, keep the stronger branch. See Codex vs Claude Code for how the engines differ.

On your machine, in Codex's own session files plus Continuum's local stores. Analytics read that history into a local ledger. Nothing is uploaded for a session to work, and a transcript only becomes public if you explicitly publish a share link.

A full HOME override breaks git, ssh, gh and node inside a worktree, which quietly ruins every long agent run. Surgical CODEX_HOME isolation separates the accounts without touching the rest of the toolchain, and an unresolvable pin fails closed rather than billing the wrong plan.

Continuum's own sessions run on hosts you own: your Mac, or an enrolled Linux or Windows box. You can still delegate to a Codex cloud environment from a session terminal; the workbench simply is not the thing hosting that run, and its gauges and ledger cover the local rail.

The Usage tab parses local Codex session history alongside Claude's, prices it with a LiteLLM snapshot, and rolls it up by repo, provider and day. That is the number OpenAI's account view cannot give you. Method: cost tracking.

Yes. Each managed session gets its own git worktree, and the resolver checks a rollout's recorded working directory before adopting it, so parallel runs and stray terminals cannot cross transcripts. Walkthrough: parallel Codex sessions.

Further reading

Guide

OpenAI Codex CLI: the complete guide

Install, auth, sandbox modes, config.toml, and the flags that actually change behaviour.

Read guide → Guide

Codex vs Claude Code: two terminal agents compared

Same form factor, opposite defaults on trust and autonomy. And a strong case for running both.

Read guide → Guide

How to track Claude Code token usage and cost

The JSONL files on your disk already record every token. Here is how to turn them into a number you can defend.

Read guide →
Begin

Run Codex
from Continuum.

Free app. Your subscriptions. Optional hosted inference. Mac stable - Windows and Linux desktop are beta.

Download for Mac See pricing

vendor-neutral · local-first · multi-device