Cobalt

An AI developer with its own cloud computer that writes code in your GitHub repo, runs the tests, and hands you a pull request.

Freemium· Public pricing not listed on the marketing site at time of review; a free 'Try Cobalt' onboarding is offered and users bring their own ChatGPT account (Plus/Pro) to power the Codex coding agent, so model-provider costs sit outside Cobalt.CodingAgent-agnostic wrapper: OpenAI Codex (via user's ChatGPT account), Claude Code, Cursor, OpenCode, Pi

Visit website →

Best for

Small dev teams and solo builders who want to hand off self-contained feature work, bug fixes and prototypes and get back a reviewable pull request with a live preview, without babysitting a local agent loop.

Skip if

Teams that require on-prem execution, strict data residency, self-hosted Git, or a scripted API-driven agent pipeline - Cobalt is a hosted, UI-driven SaaS today.

Cobalt (from Amaretto Software Labs) is a hosted AI coding agent that goes a step beyond chat-style assistants: every task gets its own persistent Linux computer in Cobalt's cloud, with your GitHub repository already cloned and the coding agent of your choice wired up. You describe the change you want in plain language, and the agent writes code, installs dependencies, boots the whole stack (multiple services, databases, Docker), runs the tests, and opens a pull request on your repo. A live preview of the running app appears next to the chat, so you can click through the working feature before you merge instead of eyeballing a diff. Because the work runs remotely, closing your laptop does not stop it, and you can return from any device to the same conversation, filesystem, and running preview. Cobalt is agent-agnostic: at review time it supports OpenAI Codex (using your ChatGPT account), Claude Code, Cursor, OpenCode, and Pi, and lets you pick the agent per task while keeping Cobalt's computer, preview, and Git workflow around it. Task computers can be saved as named starting points, so a repo with a long install/build boot only has to be provisioned once. Typical workflows are 'ship the small feature that keeps getting postponed', reproduce-and-fix bug tickets end-to-end, spin up several parallel experiments or refactor spikes without juggling worktrees on a laptop, and turn a rough idea into a clickable running prototype in minutes. Cobalt also connects to Azure DevOps repositories alongside GitHub, and per-repo CPU/memory sizing means heavier services get more headroom than a marketing site.

Editor's take

The 'agent with its own computer plus a live preview plus a pull request' shape is exactly what most chat-only coding assistants are missing, and Cobalt executes on it cleanly. The BYO-ChatGPT-account model keeps model costs honest, and being able to pick Codex, Claude Code or Cursor per task without leaving the workflow is genuinely useful. It is early-stage though, and the lack of published pricing plus GitHub/Azure-only reach means you should treat it as a promising side-runner rather than a Copilot replacement yet.

— The AI Tool Bible editorial team

Pros

✅ Runs a real cloud VM per task, so it can boot Docker, databases and multi-service stacks - not just a Next.js frontend.
✅ Delivers work as commits and pull requests on GitHub (or Azure DevOps) rather than pasted code snippets you still have to apply.
✅ Live in-browser preview of the running app updates as the agent edits code, so you can click through the feature before merging.
✅ Agent-agnostic: pick Codex, Claude Code, Cursor, OpenCode or Pi per task while keeping the same computer, repo and Git workflow.
✅ Long-running work continues after you close the tab and is waiting when you come back on any device.
✅ Parallel tasks each get isolated computers, so features, refactors and spikes do not fight over your local checkout.
✅ Saved computer snapshots let repos with heavy setup skip re-installing dependencies on every task.
✅ GitHub credentials are scoped to the task lifetime and are not persisted on the runner.

Cons

⚠️ No public pricing on the marketing site at time of review, and using Codex requires you to attach your own paid ChatGPT account.
⚠️ Cloud-only execution: no way to run the agent against a local checkout or an internal repo that cannot be reached from Cobalt's network.
⚠️ Only GitHub and (as of the July 2026 launch note) Azure DevOps are wired up - GitLab, Bitbucket and self-hosted Git are not first-class.
⚠️ The product is young and iterating quickly (Aug 2026 launch notes still adding core features like per-repo VM sizing), so expect churn.
⚠️ No documented public API for programmatic task orchestration - workflow is driven from Cobalt's own chat UI.
⚠️ Closed-source SaaS with your repo cloned onto a third-party VM, which is a non-starter for teams under strict data-residency or on-prem rules.

Use cases

Async feature implementation via pull requestReproduce-and-fix bug tickets end-to-endClickable prototypes from a plain-language briefParallel refactor spikes without local worktreesFull-stack tasks that need Docker and databasesTest-gated code changes with live preview reviewHanding off tasks and closing the laptopMulti-agent evaluation of the same coding task