Methodology
How we score cloud coding agents
bestcloudagents is operated by Replicas, which is one of the products we compare. That is exactly why the scoring is published here in full: the criteria, the weights, the formula and every product's answer, so you can check our work.
The formula
Each criterion has a weight. Every product gets full credit for a criterion it meets, half credit for partial support or for a claim we couldn't confirm from public docs, and nothing for a criterion it doesn't meet. A product's fit is the credit it earned divided by the total weight. Then, for each answer you marked as a must-have (for example “SOC 2 Type II today” or “run it in our own infrastructure”) that a product clearly fails, its fit is multiplied by 0.75.
If you name a preferred agent, a “Runs your agent” criterion with weight 4 is added and treated as a must-have. A platform that runs several agents gets credit for it when your preferred agent is among them.
There is no fixed winner and no fixed percentage. With some answers Replicas scores 100%; with others it comes fifth. The chooser shows the reasons for every score so you can see why.
The criteria and default weights
These weights apply before you answer anything. They reflect what most teams buying a cloud coding agent ask about first. Criteria at 0 only count when an answer turns them on.
| Criterion | What counts as a yes | Default weight |
|---|---|---|
| Multi-harness support | Runs more than one coding agent (for example Claude Code, Codex, Cursor and OpenCode) so you can pick the agent per task and switch when a better model ships. | 2 |
| Isolated cloud VM per task | Every task gets its own machine in the cloud, sharing nothing with other runs, so agents can install dependencies, run services and tests, and keep working with laptops closed. | 2 |
| Pull request output | The agent opens the pull request (or GitLab merge request) itself, ready for a human review. | 2 |
| Start from Slack | Mention the agent in a channel, thread or DM to start a task. | 1 |
| Start from Linear | Assign or mention the agent on a Linear issue to start a task. | 1 |
| Start from GitHub | Mention the agent on a GitHub issue or pull request to start or steer a task. | 1 |
| Start from GitLab | Mention the agent on a GitLab issue or merge request, and get a merge request back. | 0.5 |
| Start from Jira | A Jira issue can start a task directly (not only through a plugin the agent calls). | 0 |
| Bring your own model keys | Agents run on your own API keys or cloud accounts (Anthropic, OpenAI, AWS Bedrock, Microsoft Foundry and similar), so model spend stays on your provider bill. | 1 |
| Use your Claude or ChatGPT plan | Agents can sign in with subscriptions your engineers already pay for instead of a new model bill. | 0.5 |
| Model costs bundled in one bill | The vendor sells model usage as part of its plan, so there is one invoice and no provider accounts to manage. | 0 |
| Usage-based compute pricing | You pay per minute of compute the agents use, nothing while workspaces sleep, and no per-seat fee. | 1 |
| Nothing new to buy | Cloud runs are included in a plan the team may already have, with no separate platform to buy. | 0 |
| One vendor end to end | The same company makes the agent, the editor or CLI, and the cloud runs, so the whole team learns one product. | 0 |
| SOC 2 report | A SOC 2 report (Type I or Type II) your security team can review. | 0.5 |
| SOC 2 Type II today | A completed SOC 2 Type II attestation, not one in progress. | 0 |
| SCIM provisioning | Users are added and removed from your identity provider automatically. | 0.5 |
| Audit log | A record of who changed what across the organization. | 0.5 |
| Self-hosted or dedicated deployment | The platform can run in your own infrastructure or a single-tenant deployment, not only the vendor's shared cloud. | 0.5 |
| HIPAA BAA | The vendor signs a HIPAA Business Associate Agreement. | 0 |
| Parallel runs | Many tasks run at once, each in its own environment. | 1 |
| Scheduled and event-driven runs | Runs start on a schedule or from events (CI failures, webhooks, alerts) with nobody in the room. | 1 |
| Local IDE workflow | The agent works inside an editor or terminal on the engineer's own machine, for hands-on pairing. | 0 |
| Managed service | The vendor runs and maintains the infrastructure; you don't operate servers. | 1 |
| Open source | The tool's source is available under an open-source license, so you can run and audit it yourself. | 0 |
How your answers change the weights
How many engineers will use coding agents?
- Just me: SCIM provisioning set to 0; Audit log set to 0; Self-hosted or dedicated deployment set to 0; SOC 2 report set to 0; Local IDE workflow +1
- 2 to 10: no change
- 11 to 50: Parallel runs +0.5; Audit log +0.5; SCIM provisioning +0.5
- 51 to 200: SCIM provisioning +1; Audit log +1; SOC 2 report +1; Parallel runs +0.5
- More than 200: SCIM provisioning +1.5; Audit log +1.5; SOC 2 report +1.5; Self-hosted or dedicated deployment +1; Parallel runs +0.5
Which coding agent does your team want to use?
- Several, and we want to switch as models change: Multi-harness support +3 (must-have: Multi-harness support)
- Claude Code: Multi-harness support set to 0.5
- Codex: Multi-harness support set to 0.5
- Cursor: Multi-harness support set to 0.5
- OpenCode or open-weight models: Multi-harness support set to 0.5
- We'd rather buy one packaged agent: Multi-harness support set to 0; One vendor end to end +3
- No preference yet: Multi-harness support +1
How do your engineers want to work with agents?
- Hand off tasks and review pull requests: Isolated cloud VM per task +1; Pull request output +1.5; Managed service +0.5
- A mix of delegating and pairing: Local IDE workflow +2; Isolated cloud VM per task +0.5; Pull request output +0.5
- Mostly pairing in our own IDE or terminal: Local IDE workflow set to 6; Pull request output set to 0.5; Isolated cloud VM per task set to 0.5; Parallel runs set to 0.5; Scheduled and event-driven runs set to 0.5 (must-have: Local IDE workflow)
Do agent tasks need their own isolated VM?
- Yes, every task in its own machine: Isolated cloud VM per task +3 (must-have: Isolated cloud VM per task)
- Nice to have: Isolated cloud VM per task +1
- Not needed: Isolated cloud VM per task set to 0.25
Where should agent work start?
Answering this question first resets Start from Slack set to 0; Start from Linear set to 0; Start from GitHub set to 0.5; Start from GitLab set to 0; Start from Jira set to 0, so only the options you pick count.
- Slack: Start from Slack +2.5
- Linear: Start from Linear +2.5
- GitHub issues and PRs: Start from GitHub +2
- GitLab: Start from GitLab +2.5 (must-have: Start from GitLab)
- Jira: Start from Jira +2.5 (must-have: Start from Jira)
- Schedules, CI failures or webhooks: Scheduled and event-driven runs +2
How do you want to pay for models?
- Our own API keys or cloud accounts: Model costs bundled in one bill set to 0; Bring your own model keys +3 (must-have: Bring your own model keys)
- The Claude or ChatGPT plans we already pay for: Use your Claude or ChatGPT plan +2.5; Nothing new to buy +1
- One bill from the vendor, models included: Bring your own model keys set to 0; Use your Claude or ChatGPT plan set to 0; Model costs bundled in one bill +3
- No preference: no change
How should the platform itself charge?
- Only for compute used, nothing while idle, no seat fees: Usage-based compute pricing +2.5
- Ideally nothing new to buy: Usage-based compute pricing set to 0; Nothing new to buy +3
- Whatever is fair; price isn't the deciding factor: Usage-based compute pricing set to 0.5
What does your security review require?
Answering this question first resets SOC 2 report set to 0; SOC 2 Type II today set to 0; SCIM provisioning set to 0; Audit log set to 0; Self-hosted or dedicated deployment set to 0; HIPAA BAA set to 0, so only the options you pick count.
- A SOC 2 report: SOC 2 report +2.5 (must-have: SOC 2 report)
- SOC 2 Type II specifically, today: SOC 2 Type II today +3 (must-have: SOC 2 Type II today)
- SCIM provisioning: SCIM provisioning +2.5 (must-have: SCIM provisioning)
- Audit logs: Audit log +2.5 (must-have: Audit log)
- Run it in our own infrastructure: Self-hosted or dedicated deployment +3 (must-have: Self-hosted or dedicated deployment)
- A HIPAA BAA: HIPAA BAA +3 (must-have: HIPAA BAA)
How much will agents run at once?
- Many tasks in parallel, plus scheduled jobs: Parallel runs +2; Scheduled and event-driven runs +2
- A few at a time: Parallel runs +0.5
- One task at a time is fine: Parallel runs set to 0; Scheduled and event-driven runs set to 0
Who should run the infrastructure?
- The vendor; we don't want to run servers: Managed service +2.5
- Either is fine: no change
- Us, with open-source tooling on our own servers: Managed service set to 0; Open source +3; Self-hosted or dedicated deployment +1.5 (must-have: Open source)
Where the facts come from
Every product cell comes from that vendor's own documentation, pricing page or trust page, checked as of November 9, 2026. Sources are linked on the comparison page. We describe the hosted cloud product where one exists (Claude Code on the web, Codex cloud, Cursor Cloud Agents), because that is what a team compares when it moves agents off laptops; local CLIs and editors can behave differently.
Where a vendor's docs didn't let us confirm something, the cell says “Check” and scores half, the same as partial support. We would rather give a competitor partial credit than mark it down for something we didn't verify.
These are documentation reviews, not benchmarks. They tell you whether a product can do something, not how well its agent writes code. Run a trial with the same tasks on your finalists before you buy.
Who runs this site
bestcloudagents.com is operated by Replicas (replicas.dev), a cloud coding-agent platform. Replicas meets most of the criteria above, which is why it ranks first for many teams. It does not meet all of them: SOC 2 Type II is in progress, Jira can't start a task, and agents run only in the cloud. When those matter to you, the chooser ranks something else first.
If you think a score or a fact is wrong, tell us through the form on any page and we'll fix it.