For CTOs, VPs of Engineering and engineering leads
What is the right cloud coding agent for your team?
Answer ten questions about your team and get Replicas, Devin, Cursor, Claude Code, Codex and OpenCode ranked against your own criteria, with the reason behind every point.
Operated by Replicas, one of the platforms compared here. The criteria, the weights and every product's answers are published on How we score, and the chooser recommends another product when your answers point to one.
The eight things that decide it
Most cloud coding agents can write a decent pull request. Where they differ is which agents they run, where the work happens, how it starts, who pays for the model, and what your security team gets to see. These are the criteria the chooser scores.
Multi-harness support
Can you run Claude Code, Codex, Cursor and OpenCode, and switch when a better model ships?
Isolated cloud VMs
Does every task get its own machine, so agents can run services and tests and never share state?
Pull-request output
Does the agent open the PR, or hand you a branch to finish?
Starts where work is
Slack, Linear, GitHub and GitLab, plus schedules and webhooks.
Your own model keys
Does model spend land on your Anthropic, OpenAI, Bedrock or Foundry account, or a plan you already pay for?
Usage-based pricing
Per minute of compute, nothing while workspaces sleep, no seat fee.
Security and governance
SOC 2, SCIM, audit logs, HIPAA and self-hosting.
Parallel and scheduled runs
Many tasks at once, and work that starts with nobody in the room.
Our recommendation, and when it changes
Replicas
For teams that hand work to agents and review pull requests, Replicas meets more of these criteria than anything else we compared: six agents to choose from, an isolated Linux VM per task, pull requests opened by the agent, starts from Slack, Linear, GitHub and GitLab, your own model plans or keys, and per-minute pricing with nothing charged while workspaces sleep.
Where it falls short: SOC 2 Type II is still in progress, Jira can't start a task, and agents run only in the cloud.
- Pairing in your editor all day: Cursor, or Claude Code and Codex in the terminal.
- Jira triggers or SOC 2 Type II today: Devin or Cursor.
- One agent, on GitHub, already on Claude or ChatGPT plans: Claude Code on the web or Codex cloud, included in the plan.
- Open source end to end, on your own servers: OpenCode.
At a glance
Checked against each vendor's published docs and pricing pages as of November 9, 2026. “Check” means we couldn't confirm it from public docs; ask the vendor.
| Criterion | Replicas | Devin | Cursor | Claude Code | Codex | OpenCode |
|---|---|---|---|---|---|---|
| Agents | ||||||
| Multi-harness support | YesClaude Code, Codex, Cursor, OpenCode, Pi and Muse Code; pick the agent per task | NoDevin's own agent; you choose models inside it, not other agents | NoThe Cursor agent with a curated set of models | NoClaude Code only | NoCodex only | NoOpenCode only, but with 75+ model providers |
| Environment | ||||||
| Isolated cloud VM per task | YesEach workspace is its own isolated Linux VM (2, 4 or 8 vCPU) | YesEach session boots a fresh copy of the organization's snapshot in a Linux VM | YesCloud Agents run in isolated VMs with full development environments | YesClaude Code on the web runs each session in an Anthropic-managed Ubuntu VM (about 4 vCPU, 16 GB) | YesCodex cloud runs tasks in OpenAI VMs (2 vCPU/8 GiB on Plus, 4 vCPU/16 GiB on Pro and Business) | PartlyRuns wherever you put it; you provide and isolate the VM or container |
| Workflow | ||||||
| Pull request output | YesThe agent opens the pull request (GitHub) or merge request (GitLab) | YesReturns pull requests for review | YesCloud Agents push a branch and open pull requests with screenshots, videos and logs | PartlyCloud sessions push a branch; you create the PR from the session | PartlyYou open the PR from the finished task | Partly/oc comments through its GitHub integration; otherwise you script it |
| Local IDE workflow | NoAgents run only in the cloud; the desktop app and CLI connect to cloud workspaces and can open one in VS Code | YesDevin Desktop and Devin CLI | YesThe Cursor editor and CLI | YesTerminal CLI and IDE extensions; --teleport pulls a cloud session into your terminal | YesCLI, IDE extension and ChatGPT desktop app | YesTerminal, desktop app and IDE integrations |
| Where work starts | ||||||
| Start from Slack | YesMentions, DMs and threads | YesSlack and Microsoft Teams | Yes@cursor in Slack | CheckConfirm current Slack support with Anthropic | CheckConfirm current Slack support with OpenAI | CheckNot part of the core tool; confirm or build it |
| Start from Linear | YesAssign or mention on an issue | YesListed as an integration | Yes@cursor in Linear | CheckConfirm with Anthropic | CheckConfirm with OpenAI | CheckNot part of the core tool; confirm or build it |
| Start from GitLab | YesIssues and merge requests; one workspace can span GitHub and GitLab | YesGitLab and Bitbucket; self-managed GitLab on Enterprise | PartlyWorks on GitLab repositories; @cursor comments are documented for GitHub and Bitbucket | PartlyGitLab repos can be sent as a bundle, but cloud sessions can't push back | NoGitHub only for now; GitLab is on OpenAI's roadmap | CheckConfirm with the OpenCode project |
| Models and billing | ||||||
| Bring your own model keys | YesAnthropic, OpenAI, AWS Bedrock, Microsoft Foundry, OpenRouter, Cursor and OpenAI-compatible keys | PartlyA ChatGPT plan can pay for GPT usage; no API-key option is documented | NoCloud Agents bill your Cursor plan's usage at model rates | PartlyCloud sessions need a Claude plan; API keys and Bedrock aren't supported there (the CLI takes them) | PartlyCodex cloud requires a ChatGPT sign-in; API keys work only in the CLI | Yes75+ providers through your own keys, plus local models |
| Usage-based compute pricing | YesBilled per workspace minute by VM size; sleeping and archived workspaces cost nothing; seats are free | PartlyUsage quota plus credits; Teams adds $40 per seat | PartlyPlan usage at model rates; Cloud Agents need a Premium, Teams or Enterprise plan | NoSeat-based plans from $20/mo; no separate VM charge | NoSeat-based ChatGPT plans from $20/mo | PartlyThe software is free; your servers cost whatever you pay for them, idle or not |
| Security and governance | ||||||
| SOC 2 Type II today | NoSOC 2 Type II is in progress | YesSOC 2 Type II since September 2024 | YesSOC 2 Type II attestation | CheckCheck Anthropic's Trust Center | CheckCheck OpenAI's Trust Portal | PartlyCompliance is your infrastructure's |
| SCIM provisioning | YesSCIM provisioning on Enterprise | CheckSAML and OIDC SSO on Enterprise; confirm SCIM with Cognition | YesSSO and SCIM | CheckCheck the Claude Enterprise plan | CheckCheck the ChatGPT Enterprise plan | NoNo hosted user management |
| Self-hosted or dedicated deployment | YesSelf-hosted or dedicated single-tenant deployment on Enterprise | PartlyA dedicated VPC run by Cognition; Outposts move command execution to your machines | PartlySelf-hosted machines for tool execution; the service itself is Cursor-hosted | PartlyCloud sessions are Anthropic-hosted; the CLI can run on servers you control | PartlyCodex cloud is OpenAI-hosted; the Apache-2.0 CLI runs anywhere | YesRuns entirely on your own machines |
From the blog
The best cloud coding agents for engineering teams in 2026
Replicas, Devin, Cursor Cloud Agents, Claude Code on the web, Codex cloud and OpenCode compared on what matters to a team lead, with the cases where each one wins.
How to run Claude Code in the cloud (and keep it running when your laptop sleeps)
Four ways to run Claude Code in the cloud, from Anthropic's own Claude Code on the web to a rented VM, GitHub Actions and Replicas, with what each needs and costs.
Why your team needs a multi-harness cloud coding agent platform
What a multi-harness platform is, why the agent is becoming a commodity while the environment is not, and what to check before you commit to one agent's cloud.
Isolated VM environments for cloud coding agents, explained
Why cloud coding agents should get a fresh, isolated VM per task, how that compares with containers and shared machines, and what a good agent environment includes.
Get a second opinion on your shortlist
Tell us what you're weighing and an engineer from the Replicas team, who operate this site, will reply by email. We'll point you to whichever option fits, including when that isn't Replicas.