# Cloud agent alternatives: the open-source one you can host yourself > Most lists of cloud agent alternatives rank the vendor clouds and stop there. The fork that decides the next year is whether the agent runs in a product you rent or on a machine and a repo you own. Kortix is the open-source AI Operating System: your agents, skills, company memory, and connectors in one git repo you own. URL: https://cloudagentsguide.com/cloud-agent-alternatives ## Cloud agent alternatives, on the machine, the handoff and the license What each vendor documents about its own machine, handoff and license, checked in October 2026. | | Kortix (recommended) | Cursor Cloud | Railway Cloud Agents | Qoder Cloud Agents | | --- | --- | --- | --- | --- | | what it is | Open-source AI Operating System that gives every session a cloud computer | A cloud agent started from the Cursor editor, web or Slack | A persistent development VM for coding agents in a Railway project | A managed runtime for agents your own app calls | | where the machine is | Its own isolated cloud computer, one per session | An isolated VM Cursor provisions; tool calls can run on yours | A personal VM inside your Railway project environment | Qoder's sandbox, provisioned per session | | what comes back | Commits on a branch, opened as a change request you review | A merge-ready pull request with screenshots, videos and logs | Edited files on the VM disk, kept between sessions | Edits in the checkout; a branch and PR pushed on request | | open source | Yes | No | No | No | | self-host | Yes, and self-host is free | Tool execution only, on machines you manage | No, the agent runs on Railway | No, managed runtime only | | how you pay | Credits from $50 a month, 100 credits per dollar | Plan usage at the selected model's rate | Compute billed while the VM is awake | Prepaid credits; sandbox billed only while running | ## Rented or owned is the fork the lists skip Cloud agent alternatives split into two fields. Cursor Cloud, Railway Cloud Agents and Qoder Cloud Agents rent you a machine. The environment and the agent definition stay on their side. Kortix, Coder Agents and OpenHands put the machine and its definition where your team can reach them. A year from now, that difference decides what a bill, an audit and a migration look like. ## What the rented clouds hand back Cursor Cloud is the shortest hop for a team already inside the editor. Its docs describe "isolated VMs in the cloud with full development environments". Each VM clones the repo, works on a separate branch and pushes changes back for handoff. Each run returns merge-ready pull requests with screenshots, videos and logs, plus a remote desktop you can take over ([cursor.com/docs/cloud-agent](https://cursor.com/docs/cloud-agent)). Cloud Agents need a Premium, Teams or Enterprise plan, and tool execution can move to hardware you manage. The agent, the environment definitions and the model menu stay Cursor's. Railway Cloud Agents give each agent a personal VM inside a Railway project environment. Sessions start from the CLI or a connected desktop app. The disk lasts for the life of the agent, and sleeping stops compute billing while the files stay put. ([docs.railway.com/cloud-agents](https://docs.railway.com/cloud-agents)). It fits a team that already deploys on Railway and wants an agent working next to those services. Qoder Cloud Agents is the embeddable one. You define an agent and call it through the TypeScript, Python or Go SDK. Tool calls and output stream over SSE while each session runs in its own sandbox ([docs.qoder.com/cloud-agents/overview](https://docs.qoder.com/cloud-agents/overview)). Mount a GitHub repository and the agent edits that checkout; when the task asks, it pushes a branch and opens a pull request with the gh CLI ([GitHub repositories](https://docs.qoder.com/cloud-agents/api/sessions/resources/github-repositories)). Qoder lists the pull request as an optional step and says edits stay on the session's ephemeral disk unless the agent runs git push. Billing runs on prepaid credits for model calls plus sandbox runtime, metered while the session is running ([docs.qoder.com/cloud-agents/billing](https://docs.qoder.com/cloud-agents/billing)). It suits a product team that ships an agent to its own customers. All three are good products whose machine sits in an environment you do not control. That is a fair trade for a small team with no platform engineer. It becomes a different question the first time a security review asks who can read the branch, or finance asks what leaving would cost. ## The owned branch, in practice Kortix describes the machine in the repository. The project declares its sandbox image in `kortix.yaml`. A session boots a Linux machine from that image, clones the repo into `/workspace` and cuts a branch named after the session. One isolated sandbox per session. Everything the agent edits, commits and leaves behind lives on that branch, and only what it commits survives the machine. The manifest is also where the governance sits. Version 2 makes connectors, secrets and kortix_permissions deny-by-default, so an agent that names none of them gets none. Skills follow the agent file instead: an omitted skills grant leaves the agent's own `permission.skill` rule in force, and the starter agent sets `permission: allow`, so a specialist with no skills key keeps every skill. The starter's base template writes each grant out by name, and every one of them sits in the same diff. That is a different place to make the decision than a settings screen, and it survives a review like any other config change. The handoff follows from the same shape. Session work reaches `main` through a change request. Merge is default-deny for agents. A public change request in the Kortix repository shows the shape: agent-kortix authored it, a separate review session approved it, and agent-kortix merged it under the merge grant it holds there ([Kortix on GitHub](https://github.com/kortix-ai/suna/pull/10001)). Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. ## Coder Agents and OpenHands, for teams that want the loop Coder Agents is the enterprise version of the owned branch. The agent loop runs inside the Coder control plane and chat history lives in the Coder database. Tool calls reach a workspace over the connection the platform already uses. There is no SaaS component. Coder states plainly that its agent is not a wrapper around Claude Code or Codex. It works with the model provider you configure ([coder.com/docs/ai-coder/agents](https://coder.com/docs/ai-coder/agents)). A team that already runs Coder and answers to a regulator should start here. OpenHands is the smallest unit of the same idea. Its open-source browser client, Agent Canvas, runs from npm or Docker and connects to a local, self-hosted or Cloud backend. The Software Agent SDK is a Python library, and Agent Server exposes conversations over REST and WebSocket, with a Sandbox Server creating the sandboxed environments they run in ([docs.openhands.dev](https://docs.openhands.dev)). It is the right size for one repository, a script and an engineer who wants to read the loop before trusting it. Neither replaces a vendor cloud with a nicer dashboard, and neither tries to. ## Which one fits, by situation **Your output is code a person reviews, on infrastructure you choose.** Kortix. The repo describes the machine, the work arrives as a change request, and the same stack runs on your own network when the review says it must. **Your team lives in the Cursor editor.** Cursor Cloud. Keeping the same agent, the same models and the same settings on both sides of the laptop saves more than a self-hosted setup would return. **The agent has to sit next to services you already deploy.** Railway Cloud Agents. It is the shortest path from a Railway project to an agent with a persistent disk in that environment. **You are shipping an agent inside your product.** Qoder Cloud Agents. The SDK, the SSE stream and the per-session sandbox are built for a service your customers call, not for a developer at a terminal. **You have a Coder deployment or one repository and a script.** Coder Agents for the first, OpenHands for the second. Both keep the loop on your side, and both ask you to run the model account yourself. ## Moving off a rented cloud without a rewrite Start with what the vendor actually holds: the environment definition, the agent prompt, the connector list and the secrets. Copy the environment into a Dockerfile or a manifest block. Move the prompt into `agents/.md` in your repo. List the connector and secret names in `kortix.yaml` without the values. Then run one real task in both places, on a repository that will not hurt, and compare the two diffs rather than the two transcripts. Pick one trigger or scheduled job, move it, and watch it for a month before the rest follows. A cloud agent that reports to a channel fails differently from one that commits a file. ## The manifest a Kortix session boots from `kortix.yaml` ```yaml # Kortix project manifest. Source of truth for project-wide config that # lives in git alongside the code. `kortix_version` pins the schema so # the platform can evolve safely. v2 manifests are YAML-only, with one # `agents:` map per project — GOVERNANCE ONLY (connectors/secrets/ # kortix_permissions/skills/workspace/enabled), plus `file`, the path of the # agent's `.md`. Agent BEHAVIOR (description, model, mode, temperature, # permission, the prompt itself, …) lives entirely in that `.md`'s # frontmatter + body — a stock OpenCode agent file works as-is. ``` Source: https://github.com/kortix-ai/suna/blob/dev/packages/starter/templates/base/kortix.yaml The base template keeps governance in the manifest and behaviour in the markdown file beside it, and both travel with the repository. ## Questions behind the cloud agent alternatives search The nearest alternative to each product, and what stays on your side. ### What is the best alternative to Cursor Cloud Agents? Kortix, when the work has to land in a repository your team owns. Each session gets its own cloud computer and returns a change request your team reviews. Stay with Cursor Cloud if your team works inside the Cursor editor every day. ### Is there an open-source cloud agent? Yes. Kortix is an open-source AI Operating System whose sessions run on cloud computers and hand back work through a change request. Coder Agents and OpenHands are open source too, and both keep the agent loop on your own infrastructure. ### What can I use instead of a managed agent cloud? A self-hosted agent. Run Kortix on your own infrastructure, put Coder Agents in an existing Coder deployment, or run OpenHands with Docker against one repository. Each one trades a vendor dashboard for a machine and a repo you maintain. ## Run a cloud agent whose setup lives in your repo Start on Kortix Cloud with your own keys, or run the same stack on your own network. [Try Kortix](https://kortix.com) · [Read the docs](https://kortix.com/docs) One isolated sandbox per session | Self-host is free | Open source. Read it, fork it, audit it.