Cloud Coding Agents (2026): Claude Code vs Codex vs Cursor

Every official cloud agent locks you into one company's agent. Claude Code cloud runs only Claude Code, and only Claude models. Codex cloud runs only Codex, and its cloud chats are locked to OpenAI models. Cursor runs its own agent on models from several labs, but it is losing OpenAI's on November 12. If you are happy inside one ecosystem, pick that company's cloud and stop reading. Pick Cloudroom if you want Claude Code, Codex and Pi side by side, on your own subscriptions, in an open-source app.

TL;DR

• Best if you want Claude and OpenAI models side by side on your own subscriptions: Cloudroom.
• Best if you only use Claude: Claude Code cloud sessions, on your existing plan, out of research preview since September 23.
• Cleanest app: Codex cloud, but OpenAI models only.
• Best for many repo hosts and team integrations: Cursor cloud agents, but less subsidized subscription and losing OpenAI models on Nov 12.
• Best for handing off whole tickets: Devin.
• Best open-source self-hosted option: OpenHands. MIT, free cloud tier, runs Claude Code and Codex through ACP.

Two things changed in the last five weeks. Claude Code cloud sessions went generally available on September 23, 2026, and OpenAI announced on August 28, 2026 that it will wind down its models in Cursor.

What is a cloud coding agent?

A cloud coding agent is usually the same agent you run locally, but the harness runs on a server instead of your laptop, so the work survives a closed lid. The model runs in a data center either way. What moves is the agent: the process that calls the model, runs commands and edits files.

You get three things. The work keeps going while your machine is off. You get an isolated environment per session or task. And you can hand work back and forth between cloud and local.

What is Cloudroom, and how is it different from official cloud agents?

Cloudroom is an open-source desktop app that runs Claude Code, Codex, Pi and Cursor on your Mac or in the cloud, each agent in its own cloud sandbox, on your own subscriptions with no token markup. Every official cloud runs only its own agent. Cloudroom runs the real Claude Code, Codex and Pi side by side, so you pick the model for the job and switch the day the leader changes. I wrote the longer Claude Code vs Codex comparison. The conclusion there applies here too: do not lock your workflow into one vendor.

Teleport moves a thread both ways, so the same conversation and full history, uncommitted work included, can run on your Mac or in a cloud sandbox. Cloud agents can reach back to your Mac when needed, reading files and running commands on your paired machine, and your Mac skills and instructions sync one way to the sandboxes. A thread can spawn other threads with any harness, so a Claude Code thread can fan out into Codex or Pi threads. Cloud agents can also see and control a virtual screen (computer use, shipped September 29 in v77).

The app is open source: the GUI is MIT at github.com/davidondrej/cloudroom-gui and Cloudroom Core is Apache 2.0 at github.com/davidondrej/cloudroom-core. Every official cloud app is closed source. OpenHands and Cloudroom are the two you can read.

The honest con: Cloudroom is invite-only and still in development. It is a newer product than all the others here. It is fully free and open source. Join the waitlist to get an invite code.

What are Claude Code cloud sessions?

Claude Code cloud sessions are Anthropic's official way to run Claude Code on an isolated Anthropic VM instead of your laptop, and they have been out of research preview since September 23, 2026. They are the pick if you only use Claude, because they cost nothing on top of a plan you already pay for.

They come with your Claude plan and share its rate limits, with no separate compute charge. Pro is $20 a month ($17 annual). Each session runs on a VM with 4 vCPUs, 16 GB of RAM and a 30 GB disk on Ubuntu 24.04. At GA, Anthropic gave existing individual Pro and Max subscribers a one-time cloud credit: $100 on Pro, $250 on Max. You have to claim it by October 7, and it expires November 4.

I used them, and they work for the basic loop: point a task at a GitHub repo, Claude works in the cloud, you review the branch. The two gaps are the model and the fan-out. Models are Claude only, which is narrower than Cursor, never mind Cloudroom. And fan-out is weak: a normal cloud session can't start other cloud sessions. Agent teams run inside the one session and end with the task.

Projects is Anthropic's answer, now in public beta on Pro and Max and rolling out gradually, not on Team or Enterprise yet. A Project is one coordinating conversation that starts a cloud thread per task. You can turn an existing session into one with Continue as a project. I have not used it. It feels like a copout to me: it's a separate beta product, code has to be on github.com, every thread runs on Claude, and Anthropic's own docs say a project burns your limits faster than a single session.

Local setup does not carry over. Your user CLAUDE.md, user skills and user MCP servers stay on your machine (what carries over is listed here). Repo access is GitHub, with other hosts only via a one-way bundle that cannot push back. You cannot send a running terminal session to the cloud. --cloud starts a new one, the Desktop "Continue in" pushes a branch and summary (clean tree required), and --teleport pulls a cloud session down to your terminal. If you want Claude Code steerable from your phone or laptop, read the Claude Code Remote Control guide.

What is Codex cloud?

Codex cloud runs the Codex agent in an OpenAI-managed container, and it has the cleanest app of the three I use. If you live in OpenAI's world, it is a good cloud.

The catch is the model ceiling. Codex cloud runs OpenAI models only, so you cannot get a second opinion from another lab's model inside Codex cloud.

It uses your ChatGPT plan, sharing the 5-hour window and weekly cap, and OpenAI warns that cloud chats use more of your allowance than local messages. Plus is $20 a month, Pro from $100. Each task gets a container, a setup script can reach the internet while the agent phase stays offline by default, secrets are removed before the agent starts, and containers are cached for up to 12 hours. Repos come from GitHub or GitLab (beta). codex cloud submits work and applies the resulting diff locally.

My real complaint is the direction of the app. It keeps adding Work features, chat features and connectors, and the app is getting confusing. The docs now live under "ChatGPT Work and Codex", with Sites, Voice and ChatGPT for Excel sharing the same usage. I want a coding agent, not a bundle. The app is closed source. For the model-by-model detail, Claude Code vs Codex covers it.

What are Cursor cloud agents?

Cursor cloud agents run the Cursor agent on isolated Cursor VMs, and they support the most repository hosts of the group: GitHub, GitLab, Bitbucket and Azure DevOps, plus integrations like Slack, Microsoft Teams, Jira and Linear. For a team that lives across hosts, that is a real win.

The two problems are the models and the bill. On models, Cursor is losing OpenAI. OpenAI posted on August 28, 2026 that it will wind down the contract providing OpenAI models to Cursor, with a proposed shutoff on November 12, 2026, and that it will not provide future models, explicitly Astra. Today Cursor still lists GPT-5.6 Sol, Terra and Luna, but no GPT-6, and that list only shrinks from here.

On the bill, Cursor's plans are subsidized, just less than the Claude and ChatGPT plans. Every plan includes a set amount of usage, with a bigger pool for Cursor's own models (Grok and Composer). Other labs' models draw from it at API prices, and cloud agents run on the same usage. Cursor's docs say daily agent users typically spend $60 to $100 a month and power users often $200 or more, and Cursor recommends Pro+ ($60 a month) for daily agent users. In my experience you get less work out of $20 in Cursor than out of the same $20 in Claude Code or Codex.

I use it, and it works. I dislike how hard it pushes Auto mode. I want to choose my own model, and all Auto modes bill at the list price of the model each request is routed to, so Cursor picks the model and you pay that model's list price. Self-hosted machines run tool calls on hardware you control through My Machines or Team Pools, but the agent loop still runs in Cursor's cloud. Pro starts at $20 a month plus usage.

What does Devin do differently?

Devin is built for handing off whole tickets. Its docs pitch it at Linear and Jira tickets, and /handoff sends a local Devin session to the cloud and pulls the finished branch back. If your unit of work is a ticket you want to walk away from, that is the product.

Devin Cloud starts at $20 a month on Pro, with Max at $200 and Teams from $80 a month, with full seats at $40. It runs up to 10 concurrent sessions on Pro and Max, on a SWE-2 model plus OpenAI, Claude, Gemini, SpaceXAI and open models. Pro includes a daily and weekly usage quota across Devin Cloud, the CLI and Desktop, and extra usage is billed at API prices. It's pretty generous, still less than you get from a Claude or ChatGPT plan. Sessions sleep after 30 idle minutes. There is a handoff plugin for Claude Code, Codex and Cursor, so you can throw a task from your current agent to Devin without leaving your tool.

Devin is its own agent. You are not running Claude Code or Codex, you are running Devin, on Devin's quota. For ticket handoff that is fine. For people who want to keep their own harness and models, it does not replace the official clouds or Cloudroom.

Is OpenHands the best open-source cloud coding agent?

OpenHands is the open-source option, MIT licensed, with a free individual cloud tier of 10 conversations a day. Its own agent runs on models you bring by API key or at cost. It also runs Claude Code, Codex and Gemini through ACP, and OpenHands says those can use your subscription login in its cloud, injected at session start.

If you want to self-host an open-source agent platform and control every key, OpenHands is the right choice. It supports GitHub, GitLab and Bitbucket. Cloudroom is open source too. The difference is the workflow around the agents: a desktop app where the same thread moves between your Mac and the cloud, and cloud agents that can reach your Mac.

What is Jules?

Jules is Google's cloud coding agent. It runs Gemini 3 Flash as the base model and Gemini 3.1 Pro as the default on the Pro plan. The free tier gives you 15 tasks a day with 3 at once, and paid Google AI plans go up to 300 tasks and 60 at once. Paid plans only work on personal @gmail.com accounts for now. It runs only Jules, on GitHub repos. I don't know anyone serious using this.

How to choose a cloud coding agent in 2026

Choose by what you are locked into. Every official cloud runs only its own agent, and your choice of model rides on that.

• You want Claude and OpenAI models side by side on your own subscriptions: Cloudroom. Claude Code, Codex and Pi, no token markup, open-source app.
• You only use Claude: Claude Code cloud sessions. On your existing plan, no compute charge.
• You only use OpenAI models and want the cleanest app: Codex cloud.
• You need many repo hosts and team integrations: Cursor cloud agents. Real win on GitLab, Bitbucket and Azure DevOps, but its plan is less subsidized than Claude's or ChatGPT's, and you are losing OpenAI's models.
• Your unit of work is a whole ticket to delegate: Devin. It is built for tickets you hand off and walk away from.
• You want to self-host an open-source agent platform: OpenHands. MIT, free tier, full control.
• You live in Google's stack: Jules.

None of the official clouds protect you when the company changes its mind. Cursor's OpenAI models show how fast that can happen.

Agents that keep working with a closed laptop are the whole point of the cloud. If you want every way to do that, read the guide to keeping coding agents running with the laptop closed. If you want the multi-model setup without the lock-in, join the Cloudroom waitlist.