Vibe Coding
Background Coding Agents Compared (2026): Codex, Jules, Copilot, Devin
Codex, Jules, Devin, Copilot coding agent, Claude Code on the web and Cursor cloud agents: how you hand off a task, where it runs, the PR flow and the price.
A background coding agent takes a task, runs it on a machine in the cloud, and hands you a branch or a pull request while you do something else. For a solo developer or a small team, the cheapest way in is Google's Jules, which has a no-cost plan, followed by GitHub Copilot Pro at $10 a month if your code already lives on GitHub. OpenAI Codex, Claude Code on the web, Cursor cloud agents and Devin are all included in plans you may already pay for.
This post covers only the asynchronous, runs-without-you kind. Editors with an agent inside and terminal agents you watch live are in vibe coding tools compared. Everything below was checked against each vendor's own page on 11 November 2026.
What counts as a background agent
Three things: you can leave once the task is handed off, the work happens on a machine you do not run, and the result comes back as something you review in git. That rules out a chat panel in your editor, however capable. It also means the questions that matter are less about the model and more about plumbing: where you hand the task off from, what the agent can reach while it works, and how its changes land in your repository.
The agents
OpenAI Codex (cloud)
Codex cloud runs each task in "its own workspace" that keeps going while your computer sleeps. You start it from the web or desktop app by choosing Work in > Cloud, from the mobile app, from the terminal, or from Slack. In Linear you assign the issue to Codex or mention @Codex in a comment, and it posts updates to the issue.
Repo access goes through your connected GitHub or GitLab account; you pick repositories when you create an environment, and the environment is also where internet access, package-manager domains and secrets are configured. Codex does not open pull requests on its own: you review the changed files and checks, then commit or open a PR when you are ready. In Linear, it posts a summary and a link, and you open the PR from there.
Codex is included in ChatGPT Free, Go, Plus, Pro, Business and Enterprise. Free has limited access for trying it out; Plus has expanded usage, and local and cloud work share one allowance, with weekly limits possible. The Codex pricing page does not print the Plus price; its FAQ mentions Pro tiers at $100 and $200 a month.
Google Jules
Jules clones your repo into a fresh virtual machine, installs
dependencies, and proposes a plan you approve before it changes any code. You start a task
in the web app by picking a repository and branch, or by adding the label jules to a
GitHub issue, which needs the Jules GitHub app on that repo. Jules reads an AGENTS.md at
the repo root if you have one.
When it finishes, Create branch pushes the changes to a branch you own, with Jules as the commit author, and you open the PR from GitHub. For issue-started tasks, Jules posts the pull request link back on the issue (running tasks). The VM has internet access, and the FAQ says Jules does not train on private repository content.
It is the most generous free option here. The usage limits page gives the base plan 15 tasks per rolling 24 hours with 3 running at once; Google AI Pro raises that to 100 and 15, and Google AI Ultra to 300 and 60. Paid Jules plans are only for personal @gmail.com accounts. Jules's own pages do not state the subscription prices, which vary by country.
GitHub Copilot coding agent
The Copilot coding agent lives inside GitHub. You start it from a repository's Agents tab, your Copilot home page, the dashboard, or the Copilot panel in an issue or pull request. It works on one branch per task and opens one pull request per task, and sessions run either in a cloud sandbox or in a GitHub Actions environment, the latter capped at 59 minutes. Compute comes out of your monthly GitHub account allowance, separate from AI credits. It only works with repositories hosted on GitHub.
Its guard rails are the most explicit of any agent here
(risks and mitigations):
only people with write access can trigger it; it pushes to a new copilot/ branch only;
Actions workflows on its PRs wait for someone with write access to click Approve and run
workflows; the person who asked for the PR cannot approve it; and internet access is
restricted by a firewall.
On the plans page, "Assign work to Copilot and it creates a pull request" is included in Pro ($10 a month), Pro+ ($39) and Max ($100), and not in Free.
Claude Code on the web
Claude Code cloud sessions run in
an isolated, Anthropic-managed VM and keep going after you close your laptop. You start one
at claude.ai/code, from the Code tab in the Claude mobile app, from the desktop app, or
from your terminal with claude --cloud "your task", and you can pull a session back into
your terminal later with claude --teleport.
GitHub access comes either from the Claude GitHub App or, for individual developers, from
/web-setup, which sends your local gh token to your Claude account. Either way the
credentials stay on Anthropic's servers and do not enter the VM; a proxy attaches them to
git operations. The default environment has limited network access. Claude pushes its work
to a branch, you review the diff and create the PR from the session, and with the GitHub
App installed you can turn on Auto-fix, which watches the PR and pushes fixes for CI
failures and review comments.
Cloud sessions are available on Pro, Max, Ultra and Team, and count toward your normal plan usage with no separate charge for the VM. Per Claude pricing, Pro is $20 a month, or $17 billed yearly; Free does not include Claude Code.
Cursor cloud agents
Cursor cloud agents, formerly background agents, run in isolated cloud VMs with your repo cloned, dependencies installed, and any secrets and startup commands you configure. You can start one from the Cursor desktop app, from cursor.com/agents, the iOS app, Slack, Linear, a GitHub or Bitbucket comment mentioning @cursor, or the API.
It is the widest on source control: an admin connects GitHub, GitLab, Bitbucket Cloud or Azure DevOps, and you need read-write access to the repo. The agent works on its own branch, opens pull requests in the repos it changed, and can attach screenshots, videos and logs.
Usage is billed against your plan at the selected model's rate, and you set a spend limit the first time. The pricing page includes cloud agents in Premium ($30 a month), Super ($100), Ultra ($200) and Teams (from $30 per user); Free and Plus do not include them.
Devin
Devin is a full agent workspace with a shell, an IDE and a
browser. You hand it work from the web app, the API, the Devin CLI (whose /handoff sends
longer tasks to cloud Devin), or by tagging it in a Slack or Microsoft Teams thread. Each
session "boots from a fresh copy of the machine image".
Access goes through a GitHub App, the Devin.ai Integration, installed on all or selected repositories, with read and write on contents, pull requests, issues, actions and workflows (GitHub integration). Devin opens PRs using your repository's PR template.
On the pricing page, Free is $0 with a light agent quota, and Pro at $20 a month is the plan that lists Devin Cloud access; Teams is $80 a month plus $40 per seat. Usage allowances refresh daily and weekly, with no fixed figures published.
Left out: Factory
Factory's pricing page lists "Cloud & local background agents" on its $20 Pro plan, but its public pages did not document how a cloud task is handed off or how its results reach a pull request, so it is not compared here.
Side by side
| Agent | Hand off from | Runs in | Result | Cheapest plan with it | Free option |
|---|---|---|---|---|---|
| OpenAI Codex | Web, desktop, mobile, terminal, Slack, Linear | Per-task cloud workspace | Diff; you open the PR | ChatGPT Free (limited) | Yes, limited |
| Google Jules | Web, jules label on a GitHub issue | Fresh VM per task | Branch; PR link on issue tasks | No-cost plan | 15 tasks/day |
| Copilot coding agent | Agents tab, issue or PR panel | Cloud sandbox or Actions | One PR per task | Copilot Pro, $10/month | No |
| Claude Code on the web | claude.ai/code, mobile, desktop, claude --cloud | Anthropic-managed VM | Branch; you create the PR | Claude Pro, $20/month | No |
| Cursor cloud agents | Editor, web, iOS, Slack, Linear, @cursor, API | Isolated cloud VM | Opens PRs | Premium, $30/month | No |
| Devin | Web, API, CLI, Slack, Teams | Fresh machine image | Opens PRs | Pro, $20/month | Free plan, cloud not listed |
Prices are monthly, in US dollars, as shown on each vendor's page on 11 November 2026.
How to choose
- You want to try the idea for free: Jules. Fifteen tasks a day is enough to learn whether you like reviewing agent work.
- Your code and issues are on GitHub and you want the agent inside it: Copilot coding agent on Pro, with the strictest built-in review rules.
- You already pay for ChatGPT: Codex cloud, and Linear if that is where your tasks live.
- You already pay for Claude and work in a terminal: Claude Code on the web, which moves tasks between your terminal and the cloud.
- You already use Cursor, or your repos are on GitLab, Bitbucket or Azure DevOps: Cursor cloud agents on Premium.
- Your team hands out work in Slack or Microsoft Teams: Devin, or Cursor and Codex, which also take Slack tasks.
The pattern across all of them: small, well-specified tasks come back as reviewable PRs, vague ones come back as long diffs. Write the task the way you would write an issue for a new teammate, including how to run the tests.
FAQ
What is the difference between a background agent and an IDE agent?
An IDE or terminal agent works on your machine while you watch and approve steps. A background agent works on a cloud machine after you hand off the task, and you review the result as a branch or pull request. Several products do both; this post covers only the cloud side.
Is there a free background coding agent?
Yes. Jules has a no-cost plan with 15 tasks per rolling 24 hours, and ChatGPT Free includes limited Codex access. Copilot Free does not include the coding agent, and Claude Code needs a paid Claude plan.
Can these agents see my secrets?
Only what you give them. Codex, Cursor and Claude Code all let you configure environment variables or secrets per environment, and Claude Code keeps git credentials outside the VM. Give the agent its own scoped secrets, never production ones.
Do they open pull requests automatically?
Copilot, Cursor and Devin open PRs. Codex and Claude Code push changes you turn into a PR yourself, and Jules creates a branch, posting a PR link when the task started from an issue.
For the in-editor and terminal side of the same tools, see vibe coding tools compared. If you run Claude Code or Codex locally on a Mac, these Mac apps help. More AI tooling is in the AI hub, and the rest of our developer tools are under Devtools.