The best cloud coding agents for engineering teams in 2026

Every serious coding agent can now write a reasonable pull request. So when an engineering team picks a cloud coding agent, the question is less "which agent is smartest this month" and more "where does the work run, how does it start, who pays for the model, and what can our security team see?"

This guide compares six options on those questions. It's written by bestcloudagents, which is operated by Replicas, one of the six. The criteria and every product's answers are published on How we score, and the chooser will rank something other than Replicas first when your answers point there.

The short answer

What to judge them on

We score on eight things, because these are where the products actually differ:

1. Multi-harness support. Can you run Claude Code, Codex, Cursor and OpenCode, and switch when a better model ships?
2. Isolated cloud VMs. Does every task get its own machine so the agent can install packages, run services and tests, and never share state with another run?
3. Pull-request output. Does the agent open the PR, or hand you a branch?
4. Where work starts. Slack, Linear, GitHub, GitLab, schedules and webhooks.
5. Your own model keys. Does model spend land on your Anthropic, OpenAI, Bedrock or Foundry account, or a plan you already pay for?
6. Pricing. Per seat, bundled tokens, or per minute of compute with nothing charged while idle?
7. Security and governance. SOC 2, SCIM, audit logs, HIPAA, self-hosting.
8. Parallel and scheduled runs. Many tasks at once, and work that starts with nobody in the room.

Replicas

Replicas is a cloud coding-agent platform. It doesn't make its own agent; it runs Claude Code, Codex, Cursor, OpenCode, Pi and Muse Code, each task in its own isolated Linux VM, and the agent opens the pull request (or GitLab merge request). Work starts from Slack, Linear, GitHub, GitLab, the dashboard, the API, or automations on a schedule or on GitHub, GitLab, Slack, Sentry and webhook events.

Model usage runs on your own Claude or ChatGPT plan or API keys (Anthropic, OpenAI, AWS Bedrock, Microsoft Foundry and others). Replicas bills compute per workspace minute by VM size, sleeping workspaces cost nothing, and seats are free. Enterprise adds SCIM, a HIPAA BAA, and self-hosted or dedicated deployment; there's an audit log of every configuration change.

Where it falls short: SOC 2 is Type I, with Type II in progress. Jira can't start a task. Agents run only in the cloud, so it isn't a tool for pairing in your editor.

Devin

Cognition's Devin is a packaged autonomous engineer: its own agent, a fresh VM per session, an editor (Devin Desktop, formerly Windsurf), a CLI, and DeepWiki for codebase documentation. It starts from Slack, Microsoft Teams, Linear, Jira, GitHub and GitLab, and has held SOC 2 Type II since September 2024.

You choose models inside Devin but not other agents, and model usage is bundled: plans include a quota, then credits at API prices. Teams is $80 a month plus $40 per seat.

Pick it when you want one vendor for everything, Jira is where work starts, or your review needs a Type II report today.

Cursor Cloud Agents

Cursor's Cloud Agents run the Cursor agent in isolated VMs, start from the editor, the web, Slack, Linear, GitHub comments or the API, and open pull requests with screenshots, videos and logs. Cursor holds SOC 2 Type II, ISO 27001 and ISO 42001, and offers SSO and SCIM. Model usage comes from your Cursor plan; Cloud Agents need Premium or higher.

Pick it when your engineers already pair with Cursor all day and want delegation in the same product.

Claude Code on the web and Codex cloud

Anthropic and OpenAI each host their own agent. Claude Code on the web runs sessions in an Anthropic VM, is included in Claude plans, works with GitHub, and pushes a branch for you to open as a PR. It doesn't take API keys or Bedrock. Codex cloud runs tasks in OpenAI VMs, is included in paid ChatGPT plans, requires a ChatGPT sign-in, and is GitHub only for now.

Pick them when you use one agent, on GitHub, and already pay for the plan. Nothing to buy, nothing to set up. More on that in how to run Claude Code in the cloud.

OpenCode, self-hosted

OpenCode is an MIT-licensed agent that works with 75+ model providers and local models. Run it on your own servers and you have complete control, plus the job of building isolation, triggers and pull-request plumbing yourself.

Pick it when open source is a requirement and a platform team wants to own the infrastructure.

How to run a fair trial

• Give the same scoped bug or feature to each finalist, from the same repository snapshot.
• Check that private packages, databases and test services actually run in the task environment.
• Read the diff, the test output and any screenshots or recordings, and time how long a human needed to get it mergeable.
• Check permissions, secrets handling, retention and deployment options for the plan you'd actually buy.
• Add up platform, compute and model cost for the trial. Bringing your own keys doesn't automatically make it cheaper.

Then run the chooser with your answers and compare its ranking to what you saw.