Cobalt Coding Agent Review: Features, Pricing, and Verdict

This Cobalt coding agent review examines a new kind of developer product: not another model or code editor, but a cloud workspace that gives existing coding agents persistent computers, repository access, application previews, browser testing, collaboration, and automation. Cobalt 1.0 became generally available on September 10, 2026, after a short public beta.

The idea is compelling for teams juggling Codex, Claude Code, Cursor, GitHub Copilot, Gemini CLI, and other agents. It also adds another bill, security boundary, and operational layer. This review uses current Cobalt documentation and independent research; it does not claim hands-on testing or treat vendor descriptions as proof of reliability.

Cobalt Coding Agent Review: Quick Verdict

Cobalt is worth evaluating if your problem is coordinating cloud coding work rather than finding one more code-generating model. Its strongest feature is a consistent environment around several agents: each task can keep a live checkout and dependencies, expose a running application for review, let an agent operate a browser, and carry work toward a pull request. Teams can share tasks, reuse worker roles, and trigger repeatable Loops on schedules or repository events.

The tradeoff is layered cost and trust. Cobalt charges for reserved compute and retained snapshots, while customers bring their own coding-agent accounts and pay those providers separately. Source code, prompts, command output, and development artifacts pass through an additional hosted platform. The public materials describe thoughtful credential boundaries and review surfaces, but Cobalt-specific independent performance evidence is still scarce.

What Is Cobalt 1.0?

Cobalt describes itself as a cloud-first harness for coding agents. A user connects a workspace, repository provider, personal coding-agent account, and repository. Each task receives a persistent cloud computer with a live checkout. The selected agent can install dependencies, edit code, start services, run tests, and return changes for review.

Cobalt does not replace Codex, Claude Code, Cursor, Copilot, Gemini CLI, Grok, Muse Code, OpenCode, or Pi. It supplies the environment and coordination layer around them. Version 1.0 expands the beta with nine agents, GitHub and Azure DevOps support, multi-repository tasks, shared workspaces, browser testing, pull-request workspaces, automation, and task-to-task coordination.

Cobalt Features That Matter

Persistent Computers and Agent Choice

A task can retain dependencies, databases, services, Git state, and working files while its computer is available. Suspending stops active compute charges, while snapshots preserve reusable state. Teams can change the selected agent without rebuilding the surrounding preview and review workflow.

That consistency is useful, but it does not make agents interchangeable. Providers differ in models, permissions, context handling, quotas, and terms. A 2026 study of 7,156 agent-created pull requests found that task type strongly affected acceptance and no agent led every category. It did not evaluate Cobalt, so it supports task-specific testing rather than a performance claim about this platform.

Preview, Agent Browser, and Review

The Code panel exposes files, Git status, unified diffs, and a running application Preview. Agent Browser can operate against that Preview, follow user flows, inspect failures, and produce recordings. Browser evidence helps catch problems a unit test can miss, although one successful flow does not prove the whole application works.

Teammates can continue shared tasks and keep discussion with the work. Pull-request workspaces connect changes, checks, comments, and follow-ups. Multi-repository tasks can hold related frontend and API checkouts. These features make review easier; they do not validate code automatically. Humans still need to inspect diffs and verify results at GitHub or Azure DevOps.

Loops and Worker Coordination

Loops start work on schedules or repository events. Worker roles package repeatable instructions, and Process as Code lets teams version roles and Loops in Git. Coordinators can create workers, follow progress, route results, and stop for decisions.

This turns Cobalt into an agent operations layer, where costs and mistakes can multiply. Start with one narrow trigger, define completion evidence, and require approval before merges, deployments, destructive changes, or external messages. SD's AI agent security guide provides a wider control framework.

Cobalt Pricing and Total Cost

Pricing is per workspace, with unlimited repositories and members listed on every plan. Free includes 200 monthly credits and Starter or Small computers. Plus costs $30 per month with 2,500 credits, larger profiles, Cobalt Assistant, Vault, and collaboration. Pro costs $100 with 10,000 credits and all profiles. Enterprise pricing is custom.

Credits pay for reserved compute and retained snapshots. Published rates run from 7.5 credits per hour for a 1-vCPU, 4-GB Starter computer to 60 credits for an 8-vCPU, 32-GB X-Large computer. Included credits expire; purchased credits carry forward. Suspension stops compute charges, but retained snapshots still consume credits.

The plan is not the full bill. Cobalt does not include coding-agent subscriptions, provider API credit, or model usage. Plus and Pro can also spend credits on Cobalt Assistant. Track compute, storage, provider usage, CI, and human correction time rather than comparing monthly prices alone.

Cobalt Security and Privacy

Cobalt separates workspace repository scope from personal credentials. Each operator connects their own GitHub or Azure DevOps authorization, and the documentation says the platform does not fall back to another member's identity. GitHub App installations use scoped, short-lived tokens. Vault values remain outside the repository, and runtime credentials are removed before snapshots.

These are meaningful boundaries, but users still need narrow repository access, minimal Vault groups, provider-side revocation, and careful diff review. Repository files, issues, and pull-request comments are untrusted inputs that can influence an agent.

The privacy policy says Cobalt collects source files, prompts, command output, session metadata, and related development artifacts. It says signed-in product analytics exclude prompts, source code, repository names, filenames, command output, and snapshot names. General retention language does not provide one simple period for every project artifact.

Before connecting proprietary code, ask about current subprocessors, residency, deletion, incident response, audit, and enterprise controls. Review the selected agent provider's data terms too, because repository context and task material also reach that provider.

Cobalt Limitations and Risks

Independent testing of Cobalt 1.0's completion quality, uptime, recovery behavior, and cost per accepted change is limited. General benchmarks test underlying agents, not this harness. Treat Cobalt as promising and newly general-available rather than proven across every stack.

The product also adds roles, connections, identities, credits, computers, snapshots, and policy to manage. Event-driven Loops can repeat costly or manipulated actions when triggers and permissions are loose. Cobalt's own guidance says to review diffs before publishing and not to use Preview as production hosting. Browser recordings may also contain sensitive test data.

Cobalt Alternatives

Direct Codex or Claude Code workflows are simpler when one agent and a local environment already work well; SD's Claude Code vs Codex comparison explains the tradeoffs. Google Jules is a closer alternative for asynchronous cloud tasks. GitHub Copilot fits teams centered on GitHub, Cline offers editor-based model flexibility, and OpenHands emphasizes open-source deployment.

Cobalt stands out when a team wants several agent providers behind one shared computer, preview, review, and automation layer. If persistent environments and browser QA are rare, the extra platform may not earn its cost.

Who Should Use Cobalt?

Cobalt best fits teams running asynchronous tasks across repositories, reviewers, and agent providers. Agencies and platform teams may value repeatable roles, browser evidence, and event-driven maintenance. It is weaker for offline work, local-only source policies, simple projects, or buyers who need mature independent evidence.

Pilot it with a contained fix, a small feature with browser criteria, and an investigation that does not change code. Measure completion, tests, corrections, computer hours, provider usage, and reviewer time. The useful metric is production-qualified change, not lines generated.

Conclusion

This Cobalt coding agent review finds a credible coordination platform for teams whose coding-agent problem has shifted from generating code to running, verifying, reviewing, and repeating work. Persistent computers, multi-agent choice, live Preview, Agent Browser, shared tasks, Loops, and worker coordination make Cobalt 1.0 meaningfully different from a standalone assistant.

The case is not settled by the feature list. Customers pay for Cobalt compute and storage on top of agent-provider costs, introduce another processor for source code and task data, and still carry the verification burden. Independent Cobalt-specific evidence is limited because the product has only just reached general availability.

The right next step is a controlled Free-plan pilot with real acceptance tests, narrow permissions, and complete cost tracking. Cobalt is most valuable when its shared environment and review evidence reduce operational friction; otherwise, a direct agent workflow remains simpler.